Companies need to create a product position map and correctly grasp the current situation. In the recent increase in the number of e-commerce sites, the influence of product images on customers' willingness to purchase is huge. In other words, if we can construct a product map based on customers' impressions of product images, we can expect to make the product lineup on e-commerce sites more efficient and attractive. Semantic Differential (SD) method and Multi-dimensional scaling (MDS) have been conventionally used to create product maps. However, these methods are unsuitable for analyzing a target with a large number of products, such as EC sites, because the number of product evaluation data required for the analysis increases depending on the number of products. Therefore, the authors proposed a method to construct a product image map of product images with high accuracy from a relatively small number of pairwise comparison data. Specifically, a model that estimates the pairwise comparison data by subjects is machine-learned. The learned model is used to supplement the missing pairwise comparison data to estimate the evaluation values of product images for a specific axis. The values are used to construct a map. The method of selecting a small number of data used to train the model can quickly identify the overall trend. Therefore, this study proposes an adequate data selection criterion to improve the accuracy of the map based on the concept of active learning. Experiments show that, in addition to uncertainty considerations, it is adequate to have a uniform amount of pairwise comparison data for each subject in the problem of complementing pairwise comparison data between subjects, as in the case of this study.

In recent years, recommender systems based on machine learning have become common tools on various web services. Among recommendation models, embedding methods such as Item2Vec that utilize embedded representations of items are widely used in actual applications due to their effectiveness and ease of use. By utilizing embedded representations acquired through learning the interaction between users and items, it is easy to discover similar items from the viewpoint of the user's purchasing tendencies. In contrast, with this method, only biased items are recommended, making it difficult to ensure a wide variety of recommended items. However, there is a trade-off between the diversity of recommended items and accuracy and providing diversity in recommended items while maintaining accuracy is a challenging problem. Therefore, in this study, we propose a method to expand the new evaluation metric "recommendation region"(sum of distances of recommended items from the user vector in the embedding space) without significantly reducing accuracy. Specifically, we recommend not only items that are close to the user vector in the embedding space but also items with a certain distance based on detailed observation of the positional relationships. With the proposed method, we aim to increase user satisfaction by expanding the diversity of items that the user comes into contact with in the service. Finally, we demonstrate the usefulness of our proposed method through evaluation experiments using open-source datasets.

Recently, a large number of cooking recipes have been posted and shared on the Internet. Various machine learning techniques have been proposed to analyze those recipes. Those include a method to discover alternative ingredients by obtaining distributed representations from cooking procedures and ingredient names, or a method to extract basic procedures from common features in cooking procedures. Such methods utilize the constructed semantic space to calculate the distances among cooking procedures and ingredients for recipes, and demonstrate effectiveness by evaluating similarity of recipes. Using a similar semantic space, we can analyze not only the similarities among recipes but also their diversity. Even for the same dish name, there could be a variety of recipes, depending on the contributor. The diversity of recipes varies from dish to dish. By taking this diversity into account, it is possible to perform various analyses such as extracting recipes that are suitable for each user. In this study, we propose a method to analyze the diversity of recipes using distributed representation. In addition, we apply the proposed method to the posted data on an actual recipe site and show its usefulness.

The household electricity consumption data contains information on the lifestyles of each household unit, which is useful for marketing of consumer durables and services for households. In general, however, only the main power consumption can be metered using a single smart meter for each residence, and the operating status of home appliances is unknown from the companies’ viewpoint. Therefore, there have been several attempts to estimate the electricity consumption of each home appliance using dis-aggregation technology. However, there have been cost problems in practical use for exact estimation, such as the need for additional sensors. In this study, the authors formulate a problem specialized for estimating the operation and non-operation of home appliances using observed main energy data as household attribute information for marketing purposes. They then propose a state estimation model in a snapshot using the assumption of normal distributions for the electricity consumption of home appliances. Simulation experiments in a virtual housing environment show that the proposed model has an effective rate of correct answers in practical use. Additionally, a relationship between the proposed method and the prior probability distribution of device states that serve as input for the estimation is shown. The proposed model is a low-cost method for estimating household attributes that do not require additional sensors, and is expected to be used as basic information for marketing strategies.

A novel technology named fashion intelligence system, which quantifies ambiguous expressions unique to fashion, such as “casual,” “adult-casual,” and “office-casual,” was previously proposed to support users in their understanding of fashion. However, the existing visual-semantic embedding (VSE) model, which forms the basis of the system, does not support images that are composed of multiple parts, such as those containing hair, tops, trousers, skirts, and shoes. Therefore, we propose a partial VSE (PVSE) model, which enables fine-grained learning of each part of the fashion outfit. The proposed model learns embedded representations via angular-based contrastive learning. This helps in retaining three existing practical functionalities and further enables image-retrieval tasks where changes are only made to specified parts and image-reordering tasks focusing on the specified parts. In other words, the proposed model enables five types of practical functionalities, even with a simple structure. Through qualitative and quantitative experiments, we demonstrate that the proposed model is superior to conventional models, without increasing computational complexity.

A novel technology named fashion intelligence system has been proposed to quantify ambiguous expressions unique to fashion, such as "casual,""adult-casual,"and "office-casual,"and to support users' understanding of fashion. However, the existing visual-semantic embedding (VSE) model, which is the basis of its system, does not support situations in which images are composed of multiple parts such as hair, tops, pants, skirts, and shoes. We propose partial VSE, which enables sensitive learning for each part of the fashion outfits. This enables five types of practical functionalities, particularly image-retrieval tasks in which changes are made only to the specified parts and image-reordering tasks that focus on the specified parts by the single model. Based on both the multiple unique qualitative and quantitative evaluation experiments, we show the effectiveness of the proposed model.

On business chat applications that are widely used in many offices, communications between the employees are made in each group called channel. From the viewpoint of business analytics, the activation of human resource can be realized by grasping the real situation of the office community based on the analysis of the communication represented as the graph structure by utilizing the data stored on the business chat system. Especially, it is desirable to analyze the characteristics of business communication between employees for each channel, which is a group created for specific communication. For the purpose of analyzing employees’ communication, it is desirable to apply clustering or visualizing analyses of channels and it seems to possible if many channel graphs are converted to structured data. In this study, we improve the Deep Divergence Graph Kernel(DDGK) which converts the graph data to the structured data for chatting data analysis, and propose a method for obtaining embedding representation suitable to analyze the channel graphs. The proposed method enables to construct the embedding space which represent the important structure of channel graphs. Moreover, we show the effectiveness of the proposed model by applying to an actual communication data stored on Slack at a Japanese company. In this data analysis, feature vectors are activated by the model, and channel graphs can be classified by graph structures.

Error-correcting output coding (ECOC) is a method for constructing a multi-valued classifier using a combination of given binary classifiers. ECOC can estimate the correct category by other binary classifiers even if the output of some binary classifiers is incorrect based on the framework of the coding theory. The code word table representing the combination of these binary classifiers is important in ECOC. ECOC is known to perform well experimentally on real data. However, the complexity of the classification problem makes it difficult to analyze the classification performance in detail. For this reason, theoretical analysis of ECOC has not been conducted. In this study, if a binary classifier outputs the estimated posterior probability with errors, then this binary classifier is said to be noisy. In contrast, if a binary classifier outputs the true posterior probability, then this binary classifier is said to be noiseless. For a theoretical analysis of ECOC, we discuss the optimality for the code word table with noiseless binary classifiers and the error rate for one with noisy binary classifiers. This evaluation result shows that the Hamming distance of the code word table is an important indicator.

Nowadays, various services like EC sites have been expanding on the Internet, and huge amount of browsing history data are being accumulated. It is, therefore, desiblack to take effective marketing action by analyzing users' interest in detail by using accumulated browsing history data. Generally, users narrow down their interest while browsing pages and finally purchase an item on the EC site. Hence, to improve the effectiveness of marketing measures, it is important to analyze the change points of users' interest and to timely conduct it based on those change. Conventionally, Topic Tracking Model (TTM) has been proposed as a method which can model the changes of uses’ interest over time by purchase history data. However, TTM assumes the relatively long purchase history data and doesn’t consider the situation where users change their interest in short browsing site transition of about several dozen pages. Therefore, TTM cannot estimate the changes of users’ interest when applied to the browsing history data. In this research, we propose a new topic model which estimates parameters in step by step, named Time Window Topic Model. The proposed method enables us to estimate parameters and clarify the change of users' interest while browsing. It will be possible to conduct effective marketing measure based on the convergent of users' interest by applying the proposed model. Finally, we apply the proposed method to both of the artificial and real data set and show the usefulness of our proposed method.

In this online published article, a format of few entries of Tables 4 to 7 were corrected. The Corrected tables were given below (Tables 4, 5, 6, 7): (Table presented.).

In recent years, it has become common for consumers to familiarize themselves with the latest fashion trends through the internet and engage in their own fashion-inspired shopping activities. Therefore, making fashion-inspired shopping and browsing activities (internet surfing in the fashion domain) comfortable is essential because it leads to interactions in the fashion industry. However, fashion is a fuzzy and complex domain that contains many abstract elements, and this ambiguity and complexity can hinder users’ deep interest in the fashion industry. Therefore, we define a novel technology and domain called “fashion intelligence” and propose a system based on a visual-semantic embedding method for automatically learning and interpreting fashion and obtaining answers to users’ questions. Our proposed method can embed the abundant abstract tag information in the same projective space as outfit images. Mapping of images and tags in a projective space helps search for outfit images using fashion-specific abstract words. In addition, visually estimating the degree of relevance between images and tags helps interpret abstract words. As a result, this research helps decrease fashion-specific ambiguity and complexity and supports the marketing activities and fashion choices of both experts and non-experts.

With the proliferation of E-commerce sites etc., distribution channels for second-hand fashion items are diversifying and its market is expanding. In the fashion industry, seasonal items that are in demand in particular seasons are very important. For them, excessive sales can lead to a decline in the attractiveness of the company and customer satisfaction due to subsequent shortages, but under-listing leads to the generation of bad inventory due to missed sales. To solve such trade-off problem, seasonal items need to be strategically listed by accurately predicting demand. Previously, planning and managing inventory, listing, and sales were often based on the experience and know-how of display managers. However, conventional judgments are impersonal and lack objectivity, and there are concerns that they do not solve the problem when the prediction differs from the actual. Therefore, this paper proposes a demand forecasting method that enables accurate planning of seasonal items based on more quantitative judgment. Furthermore, by applying the proposal to real data and analyzing it, we show the usefulness. In addition, we design and operate a demonstration experiment to use the method as an indicator for determining the number of new items to be listed and verify its effectiveness in real business.

Consider M-valued (Mgeq3) classification systems realized by combination of N(Ngeqlceillog_{2}Mrceil) binary classifiers. Such a construction method is called an Error Correcting Output Code (ECOC). First, focusing on a Reed-Muller (RM) code, we derive a modified RM (mRM) code to make it suitable for the ECOC. Using the mRM code and the Hadamard matrix, we introduce a simplex code which is one of the powerful equidistant codes. Next, from the viewpoint of system evaluation model, we evaluate the ECOC by using constructive coding described above. We show that they have desirable properties such as Flexible, Elastic, and Effective Elastic as M becomes large, by employing analytical formulas and experiments.

With the development of information technology, a wide variety of data have been accumulated, and there are many methods for analyzing such data. In this study, we model the input data and the metrics between the data based on the assumption that each metric is generated from a continuous latent variable. Specifically, we assume that the input data are generated using low-dimensional latent variables and their projection matrices. We describe a method for estimating the latent variables. Because the generative model defined in this study cannot obtain the Q function analytically, we use the Monte Carlo EM algorithm to approximate the Q function and investigate an efficient parameter estimation method. Experiments using artificial data and the 20 newsgroups dataset demonstrate the effectiveness of the proposed method.

A marketing policy called the”Membership Stage System” is widely used in retail business. A membership stage provides benefits to customers such as shopping points when a customer's annual cumulative purchase amount exceeds a certain threshold and the customer's stage is raised a level. As a result, the company is not only able to promote the customer's willingness to purchase, but it can also obtain the purchasing history data, thereby enabling high-quality customer analysis. The most fundamental analysis is to infer the difference of purchasing characteristics between member stages and to construct different clustering models for each member stage. However, when the clustering models are learned independently for each membership stage, it is not possible to compare the obtained clusters between membership stages. In this study, we propose a new analytical method and its learning algorithm to analyze differences in cluster distribution between membership stages. Through demonstrating the proposed model applied to an actual data set of purchasing history data on a membership stage system, the effectiveness of our proposal is clarified.

In recent years, it has become common to analyze purchase history data and take advantage of the effect on business policies. In this study, the authors focus on the case of a company that is introducing a membership stage system. Probabilistic latent semantic analysis (PLSA) is well-known as an analytical model for analysis of co-occurrence of variables in data. However, the relationship between customers and items for the customer purchase behavior of each stage based on PLSA has not shown good performance. The purchase behaviors may be slightly different between customer stages, and accordingly, the purchase behavior of customers in different stages should be represented by a similar but different model. In addition, the higher the membership stage, the fewer customers there are; therefore, it becomes difficult to accurately understand the features of the customers' purchase behavior within high membership stages. In this study, the authors propose a learning algorithm that utilizes the estimated parameters of the behavior model based on the PLSA model at a lower-stage, to estimate the parameters of models at a higher-stage to which few people belong. Moreover, numerical simulation experiments are conducted and compared to actual purchase history data to confirm the performance of the proposed method.

Purchasing actions on e-commerce sites have become very common for general consumers in recent years. Products that were used to be bought at offline shops are purchased are also handled. Such products, like gifts or durable consumer goods, are often purchased infrequently and whose prefer items change each time they are purchased. A lot of methods are proposed for analysis purchase history data in order to improve customer satisfaction. However, most of them focus on the co-occurrence relationship between customers and products and treat products purchased by the same customer as similar. Then, it is difficult to use the conventional product analysis methods that have been proposed for purchase history data is difficult for some kinds of data mentioned before. Therefore, the authors have proposed an analysis method with extracting features of products by using the coefficients of binary classifiers that discriminates product purchases or not. In this study, we conduct experiments with artificial data in order to evaluate our method. Specifically, we verify how accurately the coefficients can be estimated and under what circumstances they can be estimated more accurately.

Customer churn is considered a critical issue for all businesses, as customer loss leads to a decrease in future profits. Although acquiring new customers can address such losses, this process tends to cost more than retaining existing customers. Therefore, identifying potential churners and then retaining them is important. While churn prediction has been studied widely, current research must analyze more complex business models in response to their increase, such as platform businesses. However, modeling churn prediction for these businesses is challenging because consumer behavior over platforms is more complicated. Past approaches to churn prediction can be improved upon using recent advancements in deep learning that capture the nonlinear relationships behind the data in a data-driven manner, especially for complex business models. This study proposes a method of extracting latent features from purchase histories as explanatory variables for churn prediction using a variational autoencoder with the actual customer distribution as a prior. The proposed method is validated using real purchase data from a platform business and shows a 1.5% improvement in F-measure against the baseline, and a 20% improvement for customers with recent transactions. Subsequently, the variables are examined using several methods for data analysis to interpret the meanings of the extracted features and underlying customers’ purchasing behavior in the group of potential churners. With these analyses, the model provides practical implications for understanding what kind of purchasing behavior may lead to churn and planning effective retention strategies.

In the recent world of diverse products, it is important to identify the consumer segments that should be targeted for each product. In order to identify the consumer segments to be targeted from among all consumers, it is common practice to use target attributes such as “male office workers in their 30s.” If data including attributes, hobbies and preferences of all consumers are available, target attributes can be identified using an analytical process. However, it is not realistic for each company to collect such data from all consumers for their products because of the huge cost involved. It is therefore, common to ask a consulting company that have the web browsing and purchasing histories of various sample consumers to analyze the sample data to identify the target. In particular, clustering behavioral data such as web browsing histories of sample consumers and discovering appropriate attributes that characterize target clusters (i.e. cluster attributes) is one of the most commonly used approaches. In real situations, cluster attributes are often assigned by the analyst based on the attribute statistics of the sample consumers belonging to the cluster, and multiple cluster attributes are often assumed for each cluster. When such qualitative analysis and judgment are involved, the selection of cluster attributes strongly depends on the experience and skills of the analyst. In this study, we formulate a model for clustering sample consumers based on their web browsing history, including various interests and preferences, and assigning effective cluster attributes as targets to each cluster. In addition, we propose a method to search for the best target attributes for a given objective function. By demonstrating an analysis of an actual data set, the effectiveness of the proposed method using real data is clarified.

Visualizing social relationships by a network is useful for understanding the behavior of groups and individuals. The target of this study is a network between employees in the workplace. The construction of this network enables us to understand human relationships and managing a team. To build this network, the questionnaire and E-mail data were conventionally used. However, in this work, we use conversation history data on a chat application(Slack, etc.). We propose a method of quantifying the relationship between employees from conversation data on a chat application and visualizing it as a network between employees. Specifically, we assume that strongly related employees will make remarks at adjacent times on the chat, quantify the relationship by multivariate Hawkes process and build a network. To verify the effectiveness of the proposed model, we used Slack conversation data of a real company and extracted knowledge about team management from the network.

In recent years, explainable recommendation has been a topic of active study. This is because the branch of the machine learning field related to methodologies is enabling human understanding of the reasons for the outputs of recommender systems. The realization of explainable recommendation is widely expected to increase both user satisfaction and the demand for explainable recommendation systems. Explainable recommendation utilizes a wealth of side information (such as sellers, brands, user ages and genders, and bookmark information, among others) to expound the decision-making reasoning applied by recommendation models. In explainable recommendation, although learning side information containing numerous variables leads to rich interpretability, learning too many variables presents a challenge because decreases the amount of learning that a given computational resource can perform, and the accuracy of the recommendation model may be degraded. However, numerous and diverse variables are included in the side information stored by the actual companies operating massive real-world services. Hence, to realize practical applications of this valuable information, it is necessary to resolve problems such as computational cost. In this study, we propose a new framework for explainable recommendation based on an improved knowledge graph attention network model, which utilizes the side information of items and realizes high recommendation accuracy. The proposed framework enables direct interpretation by visualizing the reasons for the recommendations provided. Experimental results show that the proposed framework reduced computational time requirements by approximately 80%, while maintaining recommendation accuracy by enabling the model to learn the probabilistically given edges included in the graph structure. Moreover, the results show that the proposed framework exhibited richer interpretability than the conventional model. Finally, a multifaceted analysis suggests that the proposed framework is not only effective as an explainable recommendation model but also provides a powerful tool for planning various marketing strategies.

In recent years, it has become very popular to use purchase history data on e-commerce sites for marketing measures to increase sales. Under such a situation, this paper considers measures using the purchase history data of a company providing delivering services of fresh flower products through an e-commerce site. This site deals mainly with fresh flowers, and the majority of items are purchased for gifts. The demands of flower gifts are usually strongly related with certain events, such as birthday, Mother's day, opening celebration, etc. Since each customer often makes purchase only at certain event when purchasing a flower gift, and it is important to encourage them to make purchases at other events from marketing viewpoint. In addition, the appearance of fresh flowers is important, so product recommendation with product images is necessary. It is relatively easy to develop floral gifts because they consist of certain patterns such as types of fresh flowers and shapes such as bouquets. However, there is no development of product which quantitatively uses purchase history information, The purpose of this research is, therefore, to generate product images that are preferred by customers in another event, considering the characteristics of product images purchased in individual event, where it is also possible to create new product images that are not contained in existing items. The proposed model is based on Conditional Variational Auto Encoader (CVAE) and can generate image outputs by inputting product images as multi-labels of events and attributes such as age and gender of customers that greatly affect product selection. Then, after learning a generator model, we consider to analyze what kinds of new products a customer with certain attributes who purchased at certain event would newly prefer at other events by changing the labels. Furthermore, in this study, we demonstrate the validity of the model by analyzing an actual data set.

Nowadays, various services are available on the Internet, and a vast amount of website browsing history data is being accumulated. In recent years, even services with few purchase actions per user, such as booking a wedding venue or purchasing insurance, have become available on the Internet. For these services, it is assumed that the interests of users gradually change and narrow down during browsing. Then, when the users decide the product to purchase, it is considered that their interests converge on a specific subject. Therefore, it is important to implement appropriate marketing strategies depending on the degree of convergence of user interests to increase effectiveness. Therefore, a method that can analyze changing user interests over time from browsing history data is desired. In this study, we propose a Time Window Topic Model that can analyze changes in user interests by considering the interests as latent topics. The proposed method can reveal the changes in interests of users even in a real problem where it is difficult to apply conventional topic models. Finally, we verify the usefulness of the proposed method by analyzing an artificial dataset.

With the rapid development of information technology in recent years, it has become common for consumers to purchase various products via electric commerce (EC) sites. As a case study, this study focuses on ZOZOUSED, which is engaged in the business of buying used clothes from users, and reselling them as second-hand goods. From the perspective of inventory and management costs, it is desirable for items to be sold as soon as possible after they are listed on EC sites, and the number of listed items has been conventionally controlled, depending on the experience of item managers. However, owing to the subjective assessment of item managers, unnecessary price reductions for sales promotion of items, or opportunity losses triggered by an excessive number of listed items, may occur. For reasonable item management, the demand prediction for items by customers is a crucial task required to develop the optimal listing plan that balances supply and demand. Therefore, this study proposes a forecasting method of sales figures for the actual operation of listing second-hand goods, which comprises two-stage models: the first model is a prior seasonal long-term prediction of sales figures for each item group based on seasonal similarity, and the second model is a short-term fine-tuning prediction for daily operation via residual predictions with recent data. Furthermore, we apply the proposed model to the actual data of past sales figures accumulated in ZOZOUSED, and analyze the obtained results to demonstrate the usefulness of the proposed method. In addition, we empirically demonstrate the effectiveness of the proposed method by designing and performing an empirical experiment on an actual business by applying the output of the proposed method as a new index for determining the number of new items to be listed.

Job-hunting activities in Japan are different from those in other countries. The features of this are the simultaneous recruitment of new graduates, joining the company in April, and the use by most students of such resources as employment information websites. In recent years, website job boards for new graduates have provided Japanese students with assistance in finding companies for which they want to work. On these boards, students can bookmark companies that they are interested in before deciding to apply. After bookmarking, a company bookmarked by a user can examine the information again later. However, even if the students rate various companies, many of these bookmarks do not lead to job applications. In other words, this can be regarded as a lost opportunity for gaining job applications from the perspective of the company. It is important for companies to gain as many job applications as possible to be successful in their recruitment activities. Therefore, a method of analyzing the entry behavior of students on job boards using factorization machines is proposed. The model predicts whether a student will submit a job application to a company. The prediction is based on student attributes and activity information, as well as information about the companies that they are interested in, as input variables. The interactions between input variables are also considered in making the prediction. In addition, the method supports student job-hunting activities and company measures for targeting students. To clarify the proposed model, analytical experiments were conducted with actual data from a website job board for new graduates.

Data augmentation methods are used as a technique to improve generalization by increasing the number of training data in image classification. However, most of these methods are not a data driven algorithm, the degree of improvement of generalization ability by performing these data augmentation methods differs between the domains of image data for training. Generative models are researched to use for augmenting data recently. In particular, Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) that can generate clean image get attention as an excellent innovation in machine learning. As GANs extension method, there is a method called CGANs (Mirza and Osindero, 2014) that can be used for data augmentation. When enough training data for each class are not prepared for classification model, the same is true for training CGANs. In such case, CGAN generates noisy images. This makes a classification model to underfit to the original training data. Moreover, when a CGAN approximates the training data distribution, the CGAN generates new training data in the same region where training data densely exist. In such case, augmented data can't reduce overfitting on the original training data. Therefore, our research contributes to augment data which meets these two requirements. In this study, we propose a method to generate data by the class specific GAN with small training data and selectively add generated data to the training data set that improves classification accuracy by using the entropy of the classification model. The feature of the proposed method is that it focuses on the positional relationship between data and the classification hyperplane in deep learning. In the proposed method, the entropy of the classification model is used to measure the positional relationship between the classification boundary and the data. As a result, the generalization performance is improved by adding the data around the classification boundary as new training data.

<p>A general QA system realizes automatic answering by analyzing questions given by the user based on a model that retrieves appropriate answer candidates for each question from a set of documents (an answering documents retrieval model). When constructing an answering documents retrieval model for question/answer documents for a community QA (Question Answering) sites (cQA sites), a method based on a similarity measure on question documents can be a basic method. However, it is difficult to present appropriate answering documents while accurately grasping the diversity of answer documents that exist for each question as seen on a cQA site. In this study, we propose a method for constructing an answering documents retrieval model that can consider the diversity of answer documents by using a topic model. In order to verify the effectiveness of a proposed model, a verification experiment using question/answer documents actually posted on a cQA site is performed.</p>

In retail stores, there is an increasing need for predicting item demand using accumulated purchase history data to cope with the fluctuating consumer demands. These fluctuations in item demand are influenced by external factors and consumer preferences. Among these, store characteristics and weather conditions, which are closely related to consumer behavior, have strong effects on item demand. For this reason, it is very important to quantitatively grasp demand fluctuations of items that are influenced by changes in weather conditions for each store by using an integrated analysis of the purchase history data of many stores and weather conditions. In this research, we focus on the temperature difference, which is the average temperature difference from the previous day, as a weather condition affecting item sales. Because consumer feeling about a temperature is dependent on the temperature difference from the previous day, it is meaningful to construct a prediction model using this information. In this research, we propose a latent class model to express the relationship between weather conditions, store characteristics, and item demand fluctuation. Also, through an analysis experiment using an actual data set, we show the usefulness of the proposed model by extracting items that are influenced by weather conditions.

Along with recent developments of Internet society, purchasing actions on E-commerce (hereinafter called "EC") sites have become common for many consumers. On the other hand, it is known that the conversion rate (hereinafter called "CVR") on EC sites is usually several percent at most. Therefore, many EC sites desire effective measures to improve CVR. In general, a user browses several pages on an EC site before he/she decide to purchase an item and it is considered that users' intentions are reflected in their page transition tendency on an EC site. If a model analyzing the page transition data can extract users' purchasing intentions, it enables to utilize the information for making a good promotion measure. Here, it is sometimes better to assume latent classes behind the users' page transitions to understand their purchase intentions, because there are usually not only several user groups with different preferences but also plural states of purchasing intentions. However, previous models either assume the same latent topic on each page in the same session or assume a latent topic for each page every time. These models cannot handle situations where users' intentions may change during browsing but not change frequently from page to page. In this study, we propose a purchasing behavior analysis model based on Hidden Topic Markov Models (HTMM). The proposed method can divide users' browsing sequence into multiple subsequences with the same statistical characteristics according to latent topics estimated from page transitions. Then, the purchase probability of each latent topic can be obtained by using the purchase results obtained from the actual browsing history data together. By the proposed model, the purchase probabilities become possible to estimate the purchase intention of the users in real time and the information is effective for considering marketing measures. In this study, an experiment using real browsing history data is carried out and the effectiveness of the proposed method is demonstrated.

Understanding customer behaviour is crucial for business success. For achieving this goal, the Recency–Frequency–Monetary (RFM) model has been commonly recognised as an effective approach to analyse customer behaviour. However, the traditional RFM approach is a coarse method for quantifying customer loyalty and contribution that can only provide a single lump-sum value of the recency (R), frequency (F), and monetary value (M); hence, it discards information regarding customers’ product preferences. Typically, different customers make different purchases. Subsequently, purchases are likely to be different across customers. This creates data sparsity, which affects the performance of conventional clustering methods. In this study, we integrated the group RFM analysis and probabilistic latent semantic analysis models to perform customer segmentation and customer analysis. The results indicated that the developed approach takes into account the product preference and provides insight into and captures a wide variety of the types of true ordering behaviour of the company’s customers. The information allows the manager to improve customer relationships and build a personalised purchasing management system for grouping customers with similar purchasing patterns.

With the development of IoT technology, it has become possible to accumulate and regularly measure multidimensional time-series data. In this study, we focus on the usage of multidimensional time-series data from printer products' log data and propose a method for its analysis. In addition to the number of sheets printed by each customer, the log data includes various time-series information such as the amount of remaining toner, the number of stoppages that occur, and the activation times. To utilize these data for business purposes, it is desirable to construct a model for predicting future changes in use characteristics for each customer. In this study, we apply the random forest algorithm to predict such changes. However, if all measurable features of the problem are included, the model becomes complex and cannot be interpreted. Although the accuracy is relatively high if an appropriate learning algorithm is applied, the complex model tends to overfit the training data. In this paper, we propose a method to select the modeling features that can be interpreted by graph mining while maintaining accuracy. This would enable us to interpret the data at the field level and discover the hypotheses that are necessary for planned marketing policies. Finally, the proposed method is applied to real data and its efficacy is demonstrated.

Recently, restaurant guide sites providing restaurant information posted by users on the Internet have been widely used as effective tools for consumers. Users, on a restaurant guide site, utilize IDs to post their recommendation articles on restaurants, and these posted articles are a valuable information source for other users. Open users can search for restaurants and read recommendation articles posted by other users. Furthermore, they can react (e.g., “like”) to a recommendation article when they feel it is helpful or they feel like visiting the restaurant. On a target restaurant guide site, each post includes the user ID, restaurant name, recommendation sentences, etc., and the number of reactions is considered to depend on these posted contents. For users who post recommendation articles, the number of reactions to their posts represents the degree of empathy from other users and is an important motivation for posting. Therefore, posting users will benefit from guidelines on how to write good recommendation sentences to increase the number of reactions. Moreover, the number of reactions can be regarded as an important indicator of the activity level of the restaurant guide site from the viewpoint of the service operating company. Therefore, an analytical model developed using historical information such as posts and reactions by users would be useful for determining the relationship between posted contents and the number of reactions. Therefore, this paper proposes a model based on the machine learning approach to analyze the relation between the number of reactions and posted contents. Finally, we demonstrate the analysis based on the proposed model using practical data.

Recently, it has become easier to make use of various kinds of information on customers (e.g. customers' purchase history), due to the development of information technology. Especially in the marketing field, in fact, many companies try to employ customer segmentation for the services customization which leads to increase customer loyalty and to keep high customer retention. One of the well-known approaches for the customer analysis based on purchase history data is the RFM analysis. The RFM analysis is usually used to segment customers into several groups by using three variables; how long it has been since their last purchase, how many times they purchased, and how much they spent. However, the conventional method of the RFM analysis did not assume a generative model. Therefore, when applying to an actual data set and scoring each index of R, F, M scores, several problems occur. The main problem is that an analyst should arbitrarily decide the threshold for the scores of RFM. On the other hand, in the field of machine learning, the probabilistic latent semantic analysis is widely used for soft clustering. The latent class model enables us to cluster customers into latent classes and to calculate the assignment probabilities of each customer to each latent class. In this paper, we propose a new latent class model for the RFM analysis based on the purchase history data. The proposed model enables to decide the scoring of RFM and segment customers automatically, and the soft clustering approach helps the interpretation of the result. Furthermore, the proposed model takes account of the generation model of RFM scores. From the result of actual data analysis, it became clear that it is possible to extract latent classes that express the statistical characteristics of data well. Given a generative model estimated from the given data, it is also possible to predict future purchase behaviors of customers or to generate virtual data for simulation analysis and make decisions based on the result. We verify the effectiveness of our model by analyzing a real purchase history data of a Japanese major retail company.

There are serious problems with the increase of challenged persons in developing countries and we have to consider how to help them. Moreover, we also have to ask whether the Quality of Life (QOL) for them has actually improved. Do they feel despair over their real ambitions and are they content to be patient while in a state of unrealized potential? In other words, we have to have awareness not only how to deal with people who need help to live, but also how to make a society in which everyone can live independently and freely. In this research, we investigate the attitudes of Nepalese citizens who don't need a great support for living towards physically challenged persons and identify the problems with the current attitude of citizens towards them and discuss possible approaches and a vision for improving their QOL. To this end, we conducted a field survey in Nepal and Japan using a questionnaire, analyzed the data, and discussed to clarify the Nepalese awareness of physically challenged persons. In the end, we consider a possible solution to these problems from several viewpoints. By collecting data by various stratifications, it was found that there was a large difference in their awareness toward challenged persons by ethnic group, age, region, and gender as the major findings of the surveys. Also the Nepalese themselves are aware of the difference. Based on the above results, we point out that “unifying actual support and promoting citizens’ awareness,” “reviewing and strengthening the educational system,” and “strengthening the leadership of the government to eliminate the differences between countries, genders, and caste groups” are three ways to address this issue. These three issues will have a positive effect on challenged persons in Nepal and also have a positive effect on society by reducing the burden on supporters.

We have proposed an editing history visualization system which can confirm where and how the learner modified program. We utilized this system for actual flipped classroom and stored a large amount of learning logs. This learning log contains all the source code in the process of being modified until the program is completed. We developed a debugging exercise extraction system to automatically generate problems for debugging practice from this learning log. The debugging exercise extraction tool we developed extracted 18,680 source codes (which became practice problems) that included syntactic errors that could be used as a debugging exercise from 16 weeks of program edit history data (total number is 31,562 files). The execution time was 488 seconds. Since it can be analyzed only once every six months, we believe it is a sufficiently practical execution time.

It is necessary to analyze the relationships between the retail sales of various items and weather conditions. However, the relationship between the sales of each item and the weather condition may vary among stores. Additionally, it is necessary to model the statistical relationships between a wide variety of goods and weather conditions by using past sales data. In such a case, it becomes unrealistic to construct a forecast model for every individual item owing to the breadth of items and the number of retail shops. This study proposes a model to analyze the relationships between the sales of various items and weather conditions. This method can be used to decompose the data into three matrices based on the nonnegative tensor factorization (NTF) method. The results of the analysis clarified that the proposed model can identify important items whose demand is strongly influenced by weather conditions, thereby increasing the effectiveness of inventory management. Additionally, the store clusters estimated by the proposed model can facilitate the construction of regression models that demonstrate the relationship between the sales of each item and weather conditions.

To solve multiple classification problems with M (geq 3) categories, many studies have been devoted using N (geq lceillog-{2}Mrceil) binary ({0, 1}) classifiers, where these systems are known as binary Error-Correcting Output Codes (binary ECOC). As an extended version of the binary ECOC, the ternary ({0,∗1}) ECOC have also been discussed, where ternary classifiers classify data into positive examples when the element is 1, into negative examples when the element is 0, and no classification when the element is. In this paper, we discuss the ternary ECOC system from the view point of the system evaluation model based on rate-distortion function. First, we discuss a table of M code words with length N which is given by a ternary matrix W of M rows and N columns. Next, by leveraging the benchmark data for multiclass document classification which is widely used in Japan, the relationships between the probability of classification error Pe and the number of the ternary classifiers N for a given M are experimentally investigated. In addition, by assuming the M-dimensional Normal distribution for a classification data model, the relationship between Pe and N for a given M is also examined. Finally, we show by the system evaluation model that the ternary ECOC systems have desirable properties such as 'Flexible', 'Elastic', and 'Effective Elastic', when M becomes large.

Currently, most university students in Japan use Internet portal sites for job-hunting activities. However, job-hunting activities are sometimes prolonged owing to a mismatch between a student and the company requirements. To solve this problem, it is important to find the students who may not be able to finish job-hunting early; this goal can be achieved by utilizing user behavior log data stored on an Internet portal site. This study proposes appropriate statistical model based on a latent class model. Specifically, we also apply clustering approach and takes account of time-series variation. The proposed model enables us to analyze entry patterns from the viewpoint of time-series variation of job-hunting activities and to predict the finish date of job-hunting for each cluster. Through the simulation experiments, the effectiveness of the proposed method was clarified. We used actual data of students' activities from an Internet portal site to demonstrate the effectiveness of the proposed method that considers the time series of the entry tendency of student users. By considering the time shift of students' preferences, it became possible to extract students who tend to struggle in job-hunting activities. It is possible to specify students who should be supported by using the proposed model.

Fast Flux Discriminant (FFD) is known as one of the high-performance nonlinear binary classifiers, and it is possible to construct a classification model considering the interaction between variables. In order to take account of the interaction between variables, FFD introduces the histogram-based kernel smoothing using subspaces including variable combinations. However, when creating a subspace, the original FFD should cover all variables including combinations of variables with low interaction. Therefore, the disadvantage is that the calculation amount increases exponentially as the dimension increases. In this study, we calculate the similarity between variables by using KL divergence. Then, among the obtained similarities, divisions are performed for each subspace with similar variables. Through this method, we try to reduce the amount of calculation while maintaining classification accuracy by using only combinations of variables that are likely to take high interaction. Through the simulation experiments with Japanese newspaper articles, the effectiveness of our proposed method is clarified.

This paper discusses the systematic application of the latent class model on business analytics. The latent class model is one of the effective statistical model classes on business analytics to represent essential statistical structures by learning the sparse and high dimensional data. This model class is useful for the purpose of reduction of feature dimension and cancellation of sparseness of the data. This is because many practical data can be assumed to consist of several unobserved subgroups. For example, a customers set, which is a target in the field of marketing analysis, consists of several subgroups with different characteristics and preferences. In addition, data clustering can also be realized by estimated belonging probabilities to latent classes of each data.This paper gives a general form of the latent class model and discuss how to construct the model structure and apply to a real problem in business analytics. After describing important points to be noted in analysis based on a latent class model, several practical examples are also shown.

In recent years, toward the realization of the Internet of Things ('IoT') society, related technologies have been developed and various electronic devices are being connected to the network. Even in companies that provide such kinds of products and services, it is possible to collect usage histories of their customers. If the companies can appropriately analyze the usage history data, it is useful for their marketing activities. However, in general, device usage histories are multivariate time-series data, and it is not obvious how to construct a feature space for customer classification and clustering. Therefore, this paper proposes a method to automatically select feature quantities characterizing the properties of customers using machine learning. We apply this method to real data and show its effectiveness.

With the development of information technology, it has become possible to collect large amounts of customer related data. In particular, retail companies have demands to extract target customers by analyzing the accumulated purchase behavior data of customers. They also need to clarify the impact of items that is related to the good customers. Therefore, a large number of retail companies use their business data for marketing activities. In this study, we focus on a company that make use of a membership-stage system. The membership-stage system sets customer ranks, which we call customer stages, based on the annual spending of each customer. Each customer is given privileges such as discounts depending on their rank. As the customer's rank move up, the customers' purchase behavior varies widely. Therefore, finding the items that lead to for customer's stage growth at each stage becomes an important goal in marketing analysis. For example, there is a framework using a linear classifier to define the importance of items between customers who belong in different stages by Yang et al.. They apply Support Vector Machine(SVM) to the purchase data of customers, classifying whether or not the customers will move up a stage in a year. During the training, moving up a stage is set positive data, and the others are set negative data. Then, they interpret the importance of items using the coefficients of the discrimination function obtained by SVM for each stage. However, the proposal of Yang et al. didn't solve the following two problems. The first problem is that the importance of items among different preferences at each stage could not be interpreted from the results. The second problem is that the features of customers that become a good customer in the first three months of the year will differ from the features of customers in the next nine month, making it difficult to devise effective feature measures. In this study, we propose an analysis method to uncover the impact of latent customer purchase behavior towards stage promotion. We only use three months of customers' purchase data to predict whether the user is capable of becoming a good customer or not using Random Forest. Here, we apply the the PLSA's results of users' purchase behavior for the explanatory variable of Random Forest and interpret the important features in of classes between good customers and the other. Moreover, we apply the proposed method to actual purchase history data to verity the performance of our proposal.

Recently, it has become popular for consumers to purchase product items through EC sites. Especially as fashion items, the purchasing actions by consumers for them through EC sites have been rapidly increased. This study focuses on a fashion EC site which operates the resale business of second-hand clothes. They assess the appropriate exhibit prices of second-hand fashion items and resell them on this EC site. A characteristic of this EC site is that if an item is not bought for a certain period, the price force to be discounted automatically. In this EC site, it is important to predict the selling price of each item in condition given information and an exhibit price. When we can predict accurate selling price and clear the effects of factors on selling price, it should help a various marketing strategies. In this paper, we propose a new regression model to predict selling price using linear regression models depending on clusters which are constructed by the relation between the features of items and seasonal off-rate. In order to show the effectiveness of our proposal, simulation experiments with a real data are demonstrated and we discuss the analysis of the results for some insightful marketing policies.

It has recently become easier for retail stores to obtain mass customer purchase history data. Analyzing these data, it is possible to understand the preferences of each customer and to use the results for marketing strategies. At the same time, it is important to take into account item seasonality in supermarkets planing marketing policies. It is, therefore, necessary to understand whether each customer purchases items based on seasonality throughout the year. In this study, we propose a new latent class model for analyzing customers’ purchasing behavior focusing on the seasonality of items, and demonstrate an analysis using our model. Moreover, we show that analysis of customers’ purchase behavior using both conventional latent class models and our latent class model provides more useful results than using only one model.

In recent years, the paradigm of social coding in software development has attracted attention to developers all over the world, and GitHub which is a social coding tool has spread to the area like education. There are many cases using it as a platform of PBL (Project Based Learning). However, since GitHub is not a tool for education, it is difficult to evaluate learners. This research focuses on the contribution of learners and proposes a system that teachers can grasp the contribution of learners.

Construction methods for multi-valued classification (multi-class) systems using binary classifiers are discussed and evaluated by a trade-off model for system evaluation based on rate-distortion theory. Suppose the multi-class systems consisted of M(≥3) categories and N(≥M-1) binary classifiers, then they can be represented by a matrix W, where the matrix W is given by a table of M code words with length N, called a code word table. For a document classification task, the relationship between the probability of classification error Pe and the number of binary classifiers N for given M is investigated, and we show that our constructed systems satisfy desirable properties such as “Flexible”, and “Elastic”. In particular, modified Reed Muller codes perform well: they are shown to be “Effective elastic”. As a second application we consider a hand-written character recognition task, and we show that the desirable properties are also satisfied.

In this manuscript, we investigate a collaborative filtering method to characterize consumption behavior of customers and services with various attributes for marketing. We assume that each customer and service have the invisible attribute which is called latent class. Assuming a combination of attribute values of a customer and service is classified to a latent class, furthermore, we propose a new Bayesian statistical model that consumption behavior is probabilistically arise based on a latent class combination of a customer, service and attribute values. Then, we show the method to estimate parameters of a statistical model based on the variational Bayes method and the mean field approximation. Consequently, we show the effectiveness of the proposed model and the estimation method by simulation.

Recently, it has become popular to purchase product items through E-commerce sites (EC sites), and the internet market scale has been expanding. Under this situation, many EC sites conduct various kinds of sales promotions by analyzing huge amount of customers' purchase histories, and customer segmentation is one of the most important tools in marketing. Particularly, modeling of customers purchase behavior based on probabilistic models such as the Aspect Model (AM) is an attractive way for customer segmentation. The AM focuses on pairs of a customer and an item and it assumes unobserved features such as customers' heterogeneity and items' similarity as latent classes. Although the original AM focuses on pairs of a customer and an item mainly, the data about customers' browsing histories are also available on EC sites. If the model can take in the information of page transitions, it becomes possible to model customers' purchase behavior in detail and make better customer segments. In this paper, we propose a new latent class model that integrates browsing histories in addition to purchase histories by assuming that customers' page transitions can be described by Markov process. An analysis of actual EC site data is demonstrated to clarify the effectiveness.

In this manuscript, we investigate a collaborative filtering method to characterize consumption behavior (or evaluation) of customers (or users) and services (or items) for marketing. Assuming that each customer and service have the invisible attribute, which is called latent class, we propose a new Bayesian statistical model that consumption behavior is probabilistically arise based on a latent class combination of a customer and service. Then, we show the method to estimate parameters of a statistical model based on the variational Bayes method and the mean field approximation. Consequently, we show the effectiveness of the proposed model and the estimation method by simulation and analyzing actual data.

Many machine learning algorithms have been proposed and applied to a wide range of prediction problems in the field of industrial management. Lately, the amount of data is increasing and machine learning algorithms with low computational costs and efficient ensemble methods are needed. Alternating Decision Forest (ADF) is an efficient ensemble method known for its high performance and low computational costs. ADFs introduce weights representing the degree of prediction accuracy for each piece of training data and randomly select attribute variables for each node. This method can effectively construct an ensemble model that can predict training data accurately while allowing each decision tree to retain different features. However, outliers can cause overfitting, and since candidates of branch conditions vary for nodes in ADFs, there is a possibility that prediction accuracy will deteriorate because the fitness of training data is highly restrained. In order to improve prediction accuracy, we focus on the prediction results for new data. That is to say, we introduce bootstrap sampling so that the algorithm can generate out-of-bag (OOB) datasets for each tree in the training phase. Additionally, we construct an effective ensemble of decision trees to improve generalization ability by considering the prediction accuracy for OOB data. To verify the effectiveness of the proposed method, we conduct simulation experiments using the UCI machine learning repository. This method provides robust and accurate predictions for datasets with many attribute variables.

Because of the advancements in information technology, a large number of document data has been accumulated on various databases and automatic multi-valued classification becomes highly relevant. This paper focuses on a multivalued classification technique that is based on Error-Correcting Output Codes (ECOC) and which combines several binary classifiers. When predicting the category of a new document data, the outputs of the binary classifiers are combined to produce a predicted value. It is a known problem that if two category sets have an imbalanced amount of training data, the prediction accuracy of a binary classifier is low. To solve this problem, a previous study proposed to employ the Reed-Muller (RM) codes in the context an ECOC approach for resolving the imbalance in the cardinality of the training data sets. However, RM codes can equalize the amount of between training data of two category sets only for a specific number of categories. We want to provide a method that can be employed for a multi-valued classification with an arbitrary number of categories. In this paper, we propose a new configuration method combining binary classifiers with categories, which are not used for classification. This method allows us to reduce the amount of training data for each binary classifier while improving the balance of the training data between two category sets for each binary classifier. As a result, the computational complexity can be decreased. We verify the effectiveness of our proposed method by conducting a document classification experiment.

With rapid development on information society, automatic document classification by machine learning has become even more important. In document classification, it is assumed that a new input data can be classified into any of the categories observed in the training data. Therefore, if a new input data belongs to an unobserved category which does not exist in the training data, then such data cannot be classified exactly. To solve the above problem, Arakawa et al. proposed the method which models the generative probabilities of documents with a mixture of Polya distributions and estimates the optimum category within all observed and unobserved categories where it is assumed that documents in each category are generated from each single Polya distribution. However, the statistical characteristics of document categories are generally more complicated and there are various underlying latent topics in a category. Because a single Polya distribution models each category in the conventional approach, this method cannot represent the variation of word frequency depending on plural unobserved latent topics. This paper proposes a new model which assumes a mixture of Polya distributions for the generative probabilities of documents in a category to represent plural latent topics. To verify the effectiveness of the proposed method, we conduct the simulation experiments of document classification by using a set of English newspaper articles.

<p>The developments in information technology have highlighted the importance of analyzing big data stored in various databases. With this as a background, the importance of distributed data mining (DDM), which is the technique of implementing data mining while databases are not transmitting raw data to each other, has been advocated. As one of the methods, Forrero et al. proposed the method of optimal learning with a support vector machine (SVM) that uses the alternating direction method of multipliers (ADMM) in the context of DDM. The apparatus is called a consensus-based distributed support vector machine (D-SVM). This method can learn the optimal hyperplane with a relatively small number of iterations and minimal communication cost for an arbitrary network structure without sharing data. However, when the statistical characteristics of data stored in each database are quite different, this method requires many iterations until convergence. Needless to say, it is better that the number of iterations and total communication cost for the learning classifier are minimized. In this study, we propose a new and effective learning method that reduces the number of iterations considering the network structure, provided that all of the nodes are connected to each other. To verify the effectiveness of the proposed method, a simulation experiment using the UCI machine learning repository and artificial data is conducted.</p>

The distance metric learning is the method to learn the distance metric from training data considering its statistical characteristics under the arbitrary constraints. To obtain the desirable distance metric, the optimization problem is solved. Most of the distance metric learning methods aim to gain the global optimal metric matrix. However there is a possibility that the global metric matrix cannot express the statistical characteristics of each category in detail. In addition, if the dimension of input data increase, the computational cost of calculating distance between data increases either. To avoid this problem, we adopt the way to use the l1 regularization to gain sparse metric matrix. By combining those, we focus on the way to deriving the plural metric matrices with a sparse structure in this study. To verify the effective ness of our proposed method, we conduct simulation experiments by using UCI machine learning repository.

The Bayes coding algorithm for the tree model class is an effective method calculating the prediction probability of appearing symbol at the next time point from the past data under the Bayes criterion. The Bayes optimal prediction is given by the mixture of all models in a given model class, and the Bayes coding algorithm gives an efficient way to calculate a coding probability. This algorithm is applicable to a general prediction problem with Time-series data. Although the Bayes coding algorithm assumes a class of Markov sources, other model classes can be useful for a real prediction problem in practice. For example, the data at the next time point may not always depend on the strict sequence of the past data. It can be possible to construct an efficient Bayes prediction algorithm for a model class on which the probability of the next symbol is conditioned by the cumulative number in a past data sequence. However, there is usually no way to previously know which model class is the best for the observed data sequence. This paper considers the method to mix the prediction probabilities given by the mixtures on different model subclass. If each calculation of the mixtures on subclasses is efficient, the proposed method is also sufficiently efficient. Based on the asymptotic analysis, we evaluate the prediction performance of the proposed method.

The use of new words, numerous spoken languages, and abbreviations on the Internet is extensive. As such, automatically acquiring words for the purpose of analyzing Internet content is very difficult. In a previous study, we proposed a method for Japanese word segmentation using character N-grams. The previously proposed method is based on a simple state-transition model that is established under the assumption that the input document is described based on four states (denoted as A, B, C, and D) specified beforehand: state A represents words (nouns, verbs, etc.)
state B represents statement separators (punctuation marks, conjunctions, etc.)
state C represents postpositions (namely, words that follow nouns)
and state D represents prepositions (namely, words that precede nouns). According to this state-transition model, based on the states applied to each pseudo-word, we search the document from beginning to end for an accessible pattern. In other words, the process of this transition detects some words during the search. In the present paper, we perform experiments based on the proposed word acquisition algorithm using Japanese and Chinese newspaper articles. These articles were obtained from Japan's Kyoto University and the Chinese People's Daily. The proposed method does not depend on the language structure. If text documents are expressed in Unicode the proposed method can, using the same algorithm, obtain words in Japanese and Chinese, which do not contain spaces between words. Hence, we demonstrate that the proposed method is language independent. &amp
copy
2016 KIIE.

In this paper, we focus on pattern recognition based on the vector space model. As one of the methods, distance metric learning is known for the learning metric matrix under the arbitrary constraint. Generally, it uses iterative optimization procedure in order to gain suitable distance structure by considering the statistical characteristics of training data. Most of the distance metric learning methods estimate suitable metric matrix from all pairs of training data. However, the computational cost is considerable if the number of training data increases in this setting. To avoid this problem, we propose the way of learning distance metric by using the each category centroid. To verify the effectiveness of proposed method, we conduct the simulation experiment by using benchmark data.

Distance metric learning is the method of learning the relevant distance metric from a training dataset by considering statistical characteristics. In order to gain a desirable distance metric, the optimization problem under an arbitrary constraint is solved. However, the representative algorithms of distance metric learning need to perform eigendecomposition at each iteration. Therefore, if the dimensions of the input data become large, the computational cost will increase drastically and it is difficult to calculate the optimal solution in a realistic amount of time. In addition, those distance metric learning methods are formulated by assuming a global (unique) metric matrix for the total vector space. Therefore, the global metric matrix cannot take into account the difference in statistical characteristics between each category. To improve those problems, the authors introduce different metric matrices for each category and propose a way to estimate the plural matrices using category information that applies the method of Mochihashi et al. The estimated metric matrices reflect the statistical characteristics of each category. The formulation of classifications by template matching and the k-NN method making effective use of the metric matrices is proposed. To verify the effectiveness of the proposed method, a simulation experiment is conducted using the benchmark data of high-dimensional and low-dimensional input data.

The electronic commerce site (EC site) has become an important marketing channel where consumers can purchase many kinds of products
their access logs, including purchase records and browsing histories, are saved in the EC sites' databases. These log data can be utilized for the purpose of web marketing. The customers who purchase many product items are good customers, whereas the other customers, who do not purchase many items, must not be good customers even if they browse many items. If the attributes of good customers and those of other customers are clarified, such information is valuable as input for making a new marketing strategy. Regarding the product items, the characteristics of good items that are bought by many users are valuable information. It is necessary to construct a method to efficiently analyze such characteristics. This paper proposes a new latent class model to analyze both purchasing and browsing histories to make latent item and user clusters. By applying the proposal, an example of data analysis on an EC site is demonstrated. Through the clusters obtained by the proposed latent class model and the classification rule by the decision tree model, new findings are extracted from the data of purchasing and browsing histories.

In this paper, we focus on classification problems based on the vector space model. As one of the methods, distance metric learning which estimates an appropriate metric matrix for classification by using the iterative optimization procedure is known as an effective method. However, the distance metric learning for high dimensional data tends to cause the problems of overfitting to a training dataset and longer computational time. In addition, the number of parameters that need to be estimated is in proportion to the square of the input data dimension. Therefore, if the dimension of input data becomes high, the number of training data to acquire a metric matrix with enough accuracy becomes enormous. Especially, these problems are caused when analyzing the document data and purchase history data stored in the EC site with high dimensional and sparse structure. To avoid these problems, we propose the method of l₁ regularized distance metric learning by introducing the alternating direction method of multiplier (ADMM) algorithm. The effectiveness of our proposed method is clarified by classification experiments using a newspaper article that has a highly dimensional and sparse structure and the UCI machine learning repository, which has a low and dense structure.

Due to the development of information technologies, there is a huge amount of text data posted on the Internet. In this study, we focus on distance metric learning, which is one of the models of machine learning. Distance metric learning is a method of estimating the metric matrix of Mahalanobis squared distance from training data under an appropriate constraint. Mochihashi et al. proposed a method which can derive the optimal metric matrix analytically. However, the vector space for document data is normally very high dimensionally and sparse. Therefore, when this method is applied to document data directly, over-fitting may occur because the number of estimated parameters is in proportion to the square of the input data dimensions. To avoid the problem of over-fitting, a regularization term is introduced in this study. The purpose of this study is to formulate the regularized estimation of the metric matrix in which the optimal metric matrix can be derived analytically. To verify the effectiveness of the proposed method, document classification using a Japanese newspaper article is conducted.

Recently, the recommender system which automatically recommends product items for customers has become more important as an efficient web marketing tool. In many electronic commerce (EC) sites, the data of customers purchases and evaluation histories are stored in a database. By using them, the system predicts users preferences, and automatically recommends items that seem to be preferred, but have not been purchased yet. In this study, we focus on the recommender system based on collaborative filtering (CF) with a latent class model. CF recommends items by using purchase or evaluation history data. Considering real purchase activity on EC sites, most of the consumers who bought items on an EC site dont post their evaluation on the site. That means more purchase history data is stored more in the database than evaluation history data. However, most studies of CF used only evaluation data to learn the model. In this case, the purchase data is not used to construct a model even though its data size is much larger than that of evaluation history. It is more desirable to learn a model by using not only evaluation history, but also purchase data to improve the CF accuracy. The purpose of this study is to construct an effective CF model to improve the CF accuracy by formulating the estimation using both evaluation history data and abundant purchase history data which has not been used in previous CF studies. Specifically, we use the aspect model, which is one of latent class models of CF. We propose a way to estimate its parameters using both evaluation history and purchase data. To verify the effectiveness of this study, a simulation experiment was conducted using a bench mark data of recommender system. We show that the prediction accuracy of the recommendation is improved.

This paper proposes a new latent class model to represent user segments in a marketing model of electric commerce sites. The aspect model proposed by T. Hofmann is well known and is also called the probabilistic latent semantic indexing (PLSI) model. Although the aspect model is one of effective models for information retrieval, it is difficult to interpret the meaning of the probability of latent class in terms of marketing models. It is desirable that the probability of latent class means the size of customer segment for the purpose of marketing research. Through this formulation, the simulation analysis to dissect the several situations become possible by using the estimated model. The impact of the strategy that we contact to the specific customer segment and make effort to increase the number of customers belonging to this segment can be predicted by using the model demonstrating the size of customer segment. This paper proposes a new model whose probability parameter of latent variable means the rate of users with the same preference in market. By applying the proposed model to the data of an internet portal site for job hunting, the effectiveness of our proposal is verified.

F. P. Preparata et al. proposed a fault diagnosis model (PMC model) to find all fault units in the multicomputer system by using outcomes that each unit tests some other units. T. Kohda proposed a highly structured(HS) system and the syndrome analysis method(SAM) to diagnose from local testing results. In this paper, we introduce the maximum a posteriori probability algorithm(MAPDA) for the HS system in the probabilistic fault model. Analyzing the MAPDA, we show that the SAM is closer to the MAPDA as the fault probability becomes smaller. Finally, we show the robustness of the SAM in the HS system.

In this paper, we focus on pattern recognition based on the vector space model with the high dimensional and sparse data. One of the pattern recognition methods is metric learning which learns a metric matrix by using the iterative optimization procedure. However most of the metric learning methods tend to cause overfitting and increasing computational time for high dimensional and sparse settings. To avoid these problems, we propose the method of l1 regularized metric learning by using the algorithm of alternating direction method of multiplier (ADMM) in the supervised setting. The effectiveness of our proposed method is clarified by classification experiments by using the Japanese newspaper article and UCI machine learning repository. And we show proposed method is the special case of the statistical sparse covariance selection.

These days, many companies store enormous amounts of text data used in their operations in digital format because computers are now used in all processes in every department. However, the valuable information in this enormous amount of text data often cannot be used effectively. Normally, such data contains a lot of useful information for company workers and it is important to use the data effectively for the development of companies. However, the volume of the text data is sometimes too enormous to use the data directly. Even if analysts spend a lot of time in order to extract useful information from the text data, it may be impossible to analyze such huge amounts of text data. Generally, clustering or grouping by similarity and naming each group to provide category information are effective ways to grasp the tendency of the whole data and systematize many ideas. However, analysis by hand can only be carried out for small volumes of data in which analysts can see all data items or ideas. Therefore, it is difficult to apply clustering by hand to the large-scale text data which is stored in companies. If clustering and naming by analysts can be applied to enormous amounts of text data, it will be useful for extracting valuable information. In this study, we propose a new method based on the combination of clustering by hand and text classification in order to effectively analyze large-scale digital data which is stored in a company. The first step of the proposed method is to provide category information by hand for the sample data selected randomly from all the text data. The next step is to estimate classifiers through learning of this sample data, and to classify the rest of the data using the classifiers automatically. Using the proposed method, enormous amounts of text data can be systemized provided that only a small sample set is analyzed by hand. To verify the effectiveness of the proposed method, it is applied to the large-scale text data which was stored in a company as a case study.

In this paper, we study a privacy preserving linear regression analysis. We propose a new protocol of a distributed calculation method that calculates a least squares estimator, in the case that two parties have different types of explanatory variables. We show the security of privacy in the proposed protocol. Because the protocol have iterative calculations, we evaluate the number of iterations via numerical experiments. Finally, we show an extended protocol that is a distributed calculation method for k parties.

The recommender system is an effective Web marketing tool that havve been used especially on electric commerce sites in recent years. The recommender system provides each user with a list of new recommended items that are predicted to be preferred by the user. Collaborative filtering is one of the most representative and powerful methods to predict user preference in the recommender system. Collaborative filtering measures the similarity of preference between users and uses it to decide items to be recommended. Based on previous researche on this method, user preference is considered to have two aspects: Purchasing interest for items and post-purchase satisfaction with items. However, the conventional methods do not consider the two different preferences at the same time. This paper suggests taking these two preferences into account and proposes a new method that allows users to choose the balance between them. The proposed method is evaluated through simulation experiments with MovieLens data. It demonstrates the effectiveness of our proposal in precision and average rating compared with a previous method.

A Bayesian network is one of the useful models for pattern recognition problems and it has the features of both stochastic prediction and causal models. A Bayesian network expresses the causal relationship between variables with directed graphs. Usually the structure of a Bayesian network is statistically estimated using a set of training data and the model selection has been applied in conventional methods when Bayesian network structures were estimated. However, it is not necessary to choose one model for the purpose of prediction. From the viewpoint of Bayesian statistics, it is well known that prediction using the mixture model on model class is Bayes optimal. In general, the mixture model that is given by a weighted sum of all models with the posterior probability on the model class is the Bayes optimal prediction. In this paper, we propose an new Bayes optimal prediction on a Bayesian network model class using the mixture model. A mixture model sometimes becomes a complex expression due to the weighted sum of all models on a model class, and it results in loss of the usefulness as a causal model. Since the easiness of interpretation is one of the merits of a Bayesian network, using a mixed model only for improvement in predictive accuracy may lead to losing the merit of a Bayesian network. Therefore, we propose a new method that is configured with the mixture model utilizing the characteristics of the Bayesian network by organizing model classes properly. Furthermore, we propose a method to quantitatively assess the strength of the causal relationship between the nodes on the mixed Bayesian network model. In addition, the effectiveness of the proposed method is clarified via a numerical experiment on an application to a prediction problem of buying and selling of shares stock market.

In this paper, we propose a method of text classification using a Bayes coding algorithm, one of the efficient data compression methods. The Bayes coding algorithm gives the Bayes optimal data compression over the tree source model class. When data is compressed by the Bayes coding algorithm, the probability structure of information sources is implicitly estimated from the compressed data. Therefore, we can expect that the implicit estimation of data compression can be utilized for other purposes, especially for the document classification problem. As for the document classification using data compression methods, ZIP format and context tree weighting methods have been proposed. However, these methods do not have Bayes optimal compression and use the compression ratio as a similarity measure between documents for classification. In the Bayes coding algorithm, a weighted mixture tree given by the compression phase can be used for estimated probability structure. Tree source is a class of Markov sources and it is possible to measure the divergence between the tree sources with the same structure. However, the Bayes coding algorithm outputs different tree structures based on the data sequence to be compressed. Since the tree structures derived from documents are different from each other, it is difficult to measure the divergence between them just as it is. This paper proposes a new method to change the structures of weighted mixture trees into the same tree structure to be able to measure the divergence. Using the divergence between trees estimated by documents, the documents can be classified. Moreover, the effectiveness of the proposed method is clarified via a simulation experiment for the document classification with natural data.

A new ensemble of rate-compatible punctured low-density paritycheck (LDPC) codes is presented. By demonstrating the code performance analytically using the Gaussian approximation, we derive a good puncturing distribution of the proposed irregular LDPC codes. From this distribution, we show that the iterative thresholds of the proposed codes for each puncturing rate are larger than those of the standard LDPC codes. We also confirm performance of the proposed LDPC code by simulations. © 2013 Pushpa Publishing House.

This paper discusses a new similarity measure between documents on a vector space model from the view point of distance metric learning. The documents are represented by points in the vector space by using the information of frequencies of words appearing in each document. The similarity measure between two different documents is useful to recognize the relationship and can be applied to classification or clustering of the data. Usually, the cosine similarity and the Euclid distance have been used in order to measure the similarity between points in the Euclidean space. However, these measures do not take the correlation among words which appear in documents into consideration on an application of the vector space model to document analysis. Generally speaking, many words which appear in documents have correlation to one another depending on the sentence structures, topics and subjects. Therefore, it is effective to build a suitable metric measure taking the correlation of words into consideration on the vector space in order to improve the performance of document classification and clustering. This paper presents a new effective method to acquire a distance measure on the document vector space based on an extended cosine measure. In addition, the way of distance metric learning is proposed to acquire the proper metric from the view point of supervised learning. The effectiveness of our proposal is clarified by simulation experiments for the text classification problems of the customer review which is posted on the web site and the newspaper article. © 2011 IEEE.

Over the past several decades the international community has identified education and capacity building as critical components in helping to shift societies toward sustainable development in a global context. However, despite calls for fostering curriculum design around education for sustainability, colleges and universities have been slow to formulate ideas and approaches to teaching. This paper introduces a collaborative educational partnership which attempts to weave in the various components connected to education for sustainability. Transformative learning theory is used together with other frameworks to guide the program and evaluate the learning outcomes of students. © 2011 Published by Elsevier Ltd.

Problems associated with document classification, an important application of text mining of text data, are focused on in this paper. There have been many models and algorithms proposed for text classification; one of these is a technique using a vector space model. In these methods, a digital document is represented as a point in the vector space which is constructed by morphological analysis and counting the frequency of each word in the document. In the vector space model, the documents can be classified using the distance measure between documents. However, there are specific characteristics in the vector space model for document classification. Firstly, it is not easy to automatically remove unnecessary words completely. The existence of unnecessary words is one of the characteristics of the text mining problems. Secondly, the dimensions of the word vector space are usually huge in comparison to the number of words appearing in a document. Although the frequencies of words appearing in a document could be small in many cases, many kinds of such words with small frequency can usually be used to classify the documents. In this paper, we evaluate the performance of document classification in the case where unnecessary words are included in the word set. Moreover, the performance of the distance measure between documents in a large dimensional word vector space is analyzed. From the asymptotic results about the distance measure, we can provide an explanation of the fact given in many experiments that classification using the empirical distance between documents calculated via the cosine measure is not particularly bad. It is also suggested that the KL-divergence is not useful for text mining problems.

Recently, it is said building customer loyalty is effective to keep the relationship between customers and companies developing. However due to this relationship's complex characteristics, there are few ways to strengthen loyalty. Consequently, it is important for company markets to grasp and calculate the important factors of customer loyalty. In order to analyze complex customer characteristics or their loyalty, one needs to analyze not only data from questionnaires, but also free-format text data. This is because the respondent can answer their opinion freely, so the collection of more information and the finding of new opinions can be expected. Recently, due to the development of information technology, people can obtain or store enormous amounts of such customer data easily. This research analyzes this free-format text data by using morphological analysis, one of the methods of natural language processing. Using this result, we have reconstructed the customer loyalty structure diagram, by using a method of analyzing customer loyalty, which was proposed by previous research. In addition to evaluating the important index of customer loyalty, we propose a method of calculation using the customer loyalty structure diagram. Moreover, this research is aiming for the evaluation of the difference between strongly satisfied and unsatisfied customers. We also investigate the difference of characteristics between product categories. From those results, corporate marketers can know what the key factors of customer loyalty are important and make a strategy to get excellent customers.

Recently, the use of affiliate sites is recognized as a way of Internet business without incurring a huge initial cost. However, there is a problem that half of the persons who are managing the affiliate sites can be earning less than 1,000 yen a month. Because this problem obstructs the development of a continuous affiliate business, it should be overcome. If the knowledge of excellent managers who are earning high profits can be used and a highly efficient way is clarified, then it will be helpful to affiliate sites operated by novices. Many interview comments and suggestions by excellent managers can be found in journals. However, an affiliate site novice cannot make use of the knowledge of excellent affiliate site managers unless it is systematic and well known. In this paper, new methods to analyze the knowledge of excellent affiliate managers and features of excellent affiliate sites are proposed. The first proposal is a technique to construct a system structure based on the techniques of excellent affiliate managers, taken from the interviews printed in journals. The system constructed here shows a list of important key points and is a systematical representation of the knowledge of excellent affiliate site managers. The second proposal is a method to clarify the features of an individual affiliate site. To do this, the items to be checked items are listed. The methodology of a useful analysis to construct and manage affiliate sites is established applying these two viewpoints. Additionally, the effectiveness of the technique proposed in this paper is clarified by applying of our proposal to a set of excellent affiliate sites.

Recently, marketing activities are gradually shifting from making new customers to keeping current customers because of the intensifying market competition. Hence, keeping customers is related to improving customer satisfaction and customer loyalty. Ultimately keeping a customer means improving of lifetime value. In addition, it has been found that there are many effects for the company. Therefore, companies aiming to improve their profits need to focus on the components of customer loyalty. However, it is difficult to specify those factors because the structure of customer loyalty is very complex and its composition differs from customer to customer. This research uses a methodology interview and free-ended questionnaire data to map the customer loyalty structure. In order to clarify constitutes customer loyalty, the method of morphological analysis as the way of natural language techniques is introduced. This is used because there is much free-ended text data that companies can accumulate easily as the result of developing information technologies. Using these data, text mining can be applied, which is one of the important ways of obtaining information for CRM(customer relationship management). Through analysis, using the concept of text mining and the introduction of knowledge of the field in the study of customer loyalty, we find the structure of loyalty and consider how to improve these assumptions. Principal component analysis using the structure extracted from the text type comments of customers and vector spaces made by the structure and word frequency in each text gives us various findings about customer loyalty.

This paper proposes a new physical distribution model to express a mixed system of direct and relay delivery methods from the manufacturer to the shop. The direct and relay delivery methods should be selected and appropriately applied depending on the situations. However, a system with the intermediate characteristics of these two methods may be effective for some settings in practice and is proposed in this paper. This method is based on an analogy of the milk-run distribution method. Milk-run logistics is a method such that the manufacturer's trucks begin at the factory, move to suppliers, collect parts at each supplier's plant, and return to the factory instead of the usual distribution by suppliers. Though milk-run logistics is an effective method as an inbound logistics system, this paper identifies how an effective distribution system can be constructed for outbound logistics. A basic system model is shown for a distribution model with three layers, including an analysis of the model by Z transform. From the numerical examples, the effectiveness of the proposal is clarified. In the proposed system, the variances of inventories in the relay distribution center and shop can be reduced.

Recently, a word-valued source has been proposed as a new class of information source models. A word-valued source is regarded as a source with a probability distribution over a word set. Although a word-valued source is a nonstationary source in general, it has been proved that an entropy rate of the source exists and the Asymptotic Equipartition Property (AEP) holds when the word set of the source is prefix-free. However, when the word set is not prefix-free (non-prefix-free), only an upper bound on the entropy density rate for an i.i.d. word-valued source has been derived so far. In this paper, we newly derive a lower bound on the entropy density rate for an i.i.d. word-valued source with a finite non-prefix-free word set. Then some numerical examples are given in order to investigate the behavior of the bounds.

In this paper, we propose a new method to analyze physical distribution models written using UML (Unified Modeling Language). Recently, the use of automated (information technology, IT) management systems for complex corporation activities and processes has been adopted by many companies. IT is, of course, a powerful tool for physical distribution processes. However, it is impossible to construct an effective system without practical knowledge. Moreover, it is sometimes difficult for system engineers constructing information systems to understand the practical knowledge and essence of activities in actual physical distribution processes. Sometimes workers are entirely ignorant when it comes to the use of information systems. From the background mentioned above, it is necessary to construct a good tool that enables both sides, the system engineers and distribution workers, to communicate and exchange information. UML is an efficient language to use for such communications. However, it is not easy to clarify complex UML models. We therefore introduce a method to analyze the structure of physical distribution described by UML and apply it to a standard logistics model. The proposal is based on multivariate analysis. Using this method, we can extract the structure of complex UML diagrams.

In this paper, we consider the Bayesian approach for representation of a set of documents. In the field of representation of a set of documents, many previous models, such as the latent semantic analysis (LSA), the probabilistic latent semantic analysis (PLSA), the Semantic Aggregate Model (SAM), the Bayesian Latent Semantic Analysis (BLSA), and so on, were proposed. In this paper, we formulate the Bayes optimal solutions for estimation of parameters and selection of the dimension of the hidden latent class in these models and analyze it's asymptotic properties.

Nishiara and Morita defined an i.i.d. word-valued source which is defined as a pair of an i.i.d. source with a countable alphabet and a function which transforms each symbol into a word over finite alphabet. They showed the asymptotic equipartition property (AEP) of the i.i.d. word-valued source and discussed the relation with source coding algorithm based on a string parsing approach. However, their model is restricted in the i.i.d. case and any universal code for a class of word-valued sources isn't discussed. In this paper, we generalize the i.i.d. word-valued source to the ergodic word-valued source which is defined by an ergodic source with a countable alphabet and a function from each symbol to a word. We show existence of entropy rate of the ergodic word-valued source and its formula. Moreover, we show the recurrence time theorem for the ergodic word-valued source with a finite alphabet. This result clarifies that Ziv-Lempel code (ZL77 code) is universal for the ergodic word-valued source.

Ziv-Lempel (ZL) codes and Bayes codes are typical universal codes. An improved algorithm of the ZL code is widely used in compression software. On the other hand, practical use of Bayes codes is difficult due to the large amount of computation needed. However, a realizable algorithm in terms of computation effort has been constructed for the FSMX model group [9]. In this paper, an information source generating a sequence by word units is assumed as a model that can represent the probabilistic structure of actual data such as text data. The asymptotic compression performance of both codes is analyzed and evaluated for the information source class (information source for the block unit) with a constant (fixed) word length. As a result, it is found that Bayes code cannot directly be universal as a coding algorithm for symbol units. On the other hand, the ZL78 code can be directly universal. Also, a configuration method for the Bayes coding method is given for an information source with a block unit.

In this paper, we discuss the difference in code lengths between the code based on the minimum description length (MDL) principle (the MDL code) and the Bayes code under the condition that the same prior distribution is assumed for both codes. It is proved that the code length of the Bayes code is smaller than that of the MDL code by o(1) or O(1) for the discrete model class and by O(1) for the parametric model class. Because we can assume the same prior for the Bayes code as for the code based on the MDL principle, it is possible to construct the Bayes code with equal or smaller code length than the code based on the MDL principle. From the viewpoint of mean code length per symbol unit (compression rate), the Bayes code is asymptotically indistinguishable from the MDL two-stage codes.

The principle of the Minimum Description Length (MDL) proposed by J. Rissanen provides a type of structure for the model estimation based on probabilistic model selection allowing minimization of the codelength. On the other hand, the use of Bayes codes makes it possible to find a coding function from a mix of probabilistic models without specifying any concrete model. It has been pointed out that codes based on the MDL principle (MDL codes) are closely related to Bayes theory because in the definition of the description length of the probabilistic model, an unknown prior distribution is assumed. In this paper, we apply asymptotic analysis to the codelength difference between the MDL codes and Bayes codes, including cases of different prior distributions. The results of the analysis clearly show that in the case of discrete model families, codes having a high prior distribution in true models (that is, the models for which an advantageous prior distribution is assumed) are favorable, but in the case of parametric model families, Bayes codes have shorter codelength than the MDL codes even in the cases of advantageous prior distribution assumed for the MDL codes. © 2000 Scripta Tech-nica.

Statistical model selection is one of the most important problems in statistics, and many works have left essential results. The conventional information criteria for model selection, such as the Akaike information criterion (AIC), the Bayesian information criterion (BIC), and minimum description length (MDL) were derived from different viewpoints. Many other model selection criteria have also been reported from various viewpoints. On the other hand, if we specify the model class and assume prior probabilities, then we can acquire Bayes optimal model selection for a finite number of samples based on Bayes decision theory. Furthermore, we can assume the various loss function adapting the purpose of model selection for practical cases. In this paper, we analyze the asymptotic properties of stasistical model selection based on Bayes decision theory. At first, we formulate Bayes optimal solution based on Bayes decision theory. In this formulation, we introduce the general loss function for practical problems. Moreover, we analyze the upper limits of the error rate of the model selection.

In this paper, we propose an asymptotic Bayes optimal prediction algorithm for linear regression model, which reduces complexity in terms of calculation. In the field of industrial engineering, linear regression analyses are mainly applied to statistical quality control and demand forecasting, owing to effectiveness of control, prediction, analysis of structure, and so on. Recently, statistical model selection has been studied as a method of estimation for linear regression models, and applied to various problems of prediction. The statistical model selection is to select a particular model out of all candidates which include the true probabilistic model. The conventional criteria for model selection are F-value, FPE and information criteria; for example, AIC, BIC, and MDL. The mainly purposes of statistical model selection are to detect the true probability model, predict for future observations, and compress the data. Since statistical model selection has many applications, it has been studied not only in the field of statistics but also in various fields of science such as information theory, automatical control theory, and so on. In the case of estimation of the linear regression model by statistical model selection, generally, a particular model is selected by information criterion based on a previous observation from all candidates. In the linear regression analysis, the model class is a set of the combination of explanatory variables. However, in the case of prediction, it is not necessary to select a particular model. IN this case, the purpose of prediction is to acquire the accuracy estimator of the future observation. Therefore, previous studies using statistical model selection for prediction may be insufficient. On the other hand, the prediction method based on Bayes decision theory has been reported in various fields. In this method, predictions using the mixture model, which is mixing all candidates, are acquired as Bayes optimal solution, which minimizes the Bayesian mean square error. For this reason, we apply the mixture model for the linear regression models for prediction. We, at first, show that prediction by the mixture model is Bayes optimal prediction. However, it is difficult to strictly calculate the mixture probability because of the integration complexity on the parameter space. Therefore, we propose a new prediction method which removes the integration on account of reducing the complexity. Strictly speaking, we propose an asymptotic Bayes optimal prediction, which calculates the asymptotic posterior predictive distribution; i.e., mixture model. At last, we verify the effectiveness of the proposal through the simulation experiments.

We analyse the extended stochastic complexity (ESC) which has been proposed by K. Yamanishi. The ESC can be applied to learning algorithms for on-line prediction and batch-learning settings. Yamanishi derived the upper bound of ESC satisfying uniformly for all data sequences and that of the asymptotic expectation of ESC. However, Yamanishi concentrates mainly on the worst case performance and the lower bound has not been derived. In this paper, we show some interesting properties of ESC which are similar to Bayesian statistics: the Bayes rule and the asymptotic normality. We then derive the asymptotic formula of ESC in the meaning of almost sure and mean convergence within an error of o(1) using these properties.

For production-inventory systems based on periodical ordering systems, it is important to control variances in order quantity and inventory. Therefore, a periodical ordering system based on optimal control theory has been proposed. The criterion of the ordering system is the weighted summation of those variances. The formula of order minimizes the criterion, and optimal order quantity can be calculated for every period. However, it is rational for the ordering system to minimize the cost function in practice. If variance in inventory is large, then a shortage or remainder of inventory would occur and inventory cost would rise. If variance in order quantity is large, then production loss and extra production costs would rise. In this paper, we formulate a periodical ordering system minimizing the cost function that depends on variances in order quantity and inventory. We show a method for searching the formula to calculate the optimal order quantity. This solution is based on optimal control theory for a linear system with colored noise.

In this paper, we consider the contingency table as a two-way classification with binomial distribution at each of the cells. The contingency table is useful for many practical cases in the field of the quality control, market research, and so on. In the analysis of data of the contingency table, the probabilistic models with parameters are assumed at each cell and the parameters are estimated from observations (data sets). In the case of binomial distribution, the parameters are the probabilities of each cell. The maximum likelihood estimator (MLE) for each cell is one of the parameter estimators. However, the precision of the MLE may not be sufficient when the sample size is small. On the other hand, if we previously assume the prior density on parameter space, we can formulate Bayes optimal estimation of the parameter with the square error loss function. This estimator is Bayes optimal for the finite sample. The another way for precise estimation is to estimate the parameter after the hypothesis, such that the parameter of some cell is equal to that of another cell. If the hypothesis is correct, the estimation error may be reduced rather than parameters being independently estimated at each cell. If we can previously assume a hypothetical set, a hypothesis may be selected from the observations, but in a true hypothesis, the set is unknown. In such a method, although we can regard a hypothesis as a model, this method can be formulated based on the statistical model selection problem, and we can apply conventional information criteria to select the model. However, in Bayes decision theory, the selection of a model is not optimal for prediction. We can formulate Bayes optimal estimator on the condition that the set of candidates of the models is given. In this paper, we propose a method to predict future observations and estimate the parameter values at each cell based on Bayes optimal solution, which uses a mixture model of all candidates. That is, we shall show that the method using the mixture model is Bayes optimal solution with the square error loss function for the parameter estimation similar to the prediction problem. Moreover, we shall show a practical algorithm to calculate Bayes optimal estimator using a mixture model assuming Bata distribution and uniform distribution as the prior distributions for the parameter space and model class, respectively. Through the simulation experiment, we shall show the properties of the proposed method for parameter estimation.

The statistical model selection problems are discussed, and the criterion based on the Bayes decision theory for the detection of the true model is derived. We propose a new Bayesian model selection scheme, which can detect the change points of the parameter of the information source and also estimate each unknown parameter. In this paper, we formulate the Bayes optimal solution of the change detection problem, and analyze its property of the consistency.

We shall generalize B.S. Clarke and A.R. Barron's analysis of the Bayes method for the FSMX sources. The FSMX source considered here is specified by the set of all states and its parameter value. At first, we show the asymptotic codelengths of individual sequences of the Bayes codes for the FSMX sources. Secondly, we show the asymptotic expected codelengths. The Bayesian posterior density and the maximum likelihood estimator satisfy asymptotic normality for the finite ergodic Markov source, and this is the key of our analysis.

With conventional models, in order to compare the properties of push-type production systems with those of pull-type systems, in previous works it was assumed that the ordering cycle period of the push-type system was equal to that of the pull-type system. The ordering cycle period of the push-type system is, however, usually different from that of the pull-type system in actual use. For example, the ordering cycles of the push and pull-type systems may be given every day and every three hours, respectively. The results of previous works with the same ordering cycle period cannot be applied to the above practical cases. In this paper we propose a new model in order to compare the properties of the push and pull-type systems under condition that the ordering cycle period may be different. The criteria in this paper are variances in inventory and order quantity. This is because we want to investigate the essence of the behaviors of the production systems. At first, we formulate optimal push and pull-type production systems. Secondly, we formulate a demand model in order to consider the differences in the ordering cycle periods, and apply this model to both systems. Finally, we evaluate the performance through numerical analysis. We analyze the rate of the ordering cycle period coinciding with the evaluated value of the push-type system with that of the pull-type system, which is denoted by τ^*. As the result, we show that the effect of τ to the variances is very strong. If autocorrelation of the demand is strong, then τ^* is large, however, if the autocorrelation is weak, then τ^* is smaller than 2.

In this paper, we shall derive the upper bounds on error rates of the statistical model selection using the information criteria, P. Similar bounds were derived by J. Suzuki (1993). We shall generalize the results for the general model class. © 1998 IEEE.

SUMMARY We shall generalize B. S. Clarke and A. R. Barren's analysis of the Bayes method for the FSMX sources. The FSMX source considered here is specified by the set of all states and its parameter value. At first, we show the asymptotic codelengths of individual sequences of the Bayes codes for the FSMX sources. Secondly, we show the asymptotic expected codelengths. The Bayesian posterior density and the maximum likelihood estimator satisfy asymptotic normality for the finite ergodic Markov source, and this is the key of our analysis.

BP (Back Propagation) learning algorithm based on Conjugate Gradient Method is an effective method for high-speed learning. This method has originally been introduced for the problem to minimize the quadratic function and is guaranteed to converge to it in the number of times equal to the search surface dimensions. It is also applied as it is to a generalized function like BP learning algorithm, because in this case too, it can approximate locally to a quadratic function. However, it fails to converge in the number of times equal to the search dimensions in this case, requiring the step to take the second approximation process after the search dimensions reached a certain number. This step is called "Restart". However, this method occasionally falls into local minimal depending on search surface, because it locally approximates a generalized function to a quadratic function. The cause is considered that it continues searching until its number of times reaches the conventionally determined number of times before the restart, even if the approximation to a quadratic function is low in precision ; therefore this restart method is considered to improve the search efficiency when the quadratic function approximation is low in precision. We therefore propose an efficient learning algorithm based on Conjugate Gradient Method. The method proposed here can decide whether the restart is needed according to new parameters which evaluate the quadratic approximation accuracy at every search point. We call the proposed method "An improved Conjugate Gradient Method".

We analyze quantitatively the the difference of codelengths between both codes for the hierarchical (nested) model class. © 1997 IEEE.

The objective of this paper is to formulate the optimal control system for the linear system with colored noise and to apply to the periodical ordering system for the push-type production and inventory systems, which provides the possible procedure to control both the variances of order quantity and inventory. This system realizes the optimization of the criterion of the weighted sum of the variances of production and inventory. For the extension to the solution of multi-stage process and of the case the demand is not stationary, it is necessary to formulate the control system by the optimal control theory. However, the optimal regulator control which is the conventional optimal control theory can be applied directly to the system with colored noise. In this paper, we propose the new formulation of this problem and the application to the production and inventory control system.

This paper proposes a new formulation which minimizes the differential entropy for an optimal control problem. The conventional criterion of the optimal regulator control is a standard quadratic cost function E[M{x(t)}(2) + N{u(t)}(2)], where x(t) is a state variable, u(t) is an input value, and M and N are positive weights. However, increasing the number of the variables of the system it is complex to find the solution of the optimal regulator control. Therefore, the simplicity of the solution is required. In contrast to the optimal regulator control, we propose the minimum entropy control which minimizes a differential entropy of the weighted sum of x(t) and u(t). This solution is derived on the assumptions that the linear control and x(t)u(t) less than or equal to 0 are satisfied. As the result, the formula of the minimum entropy control is very simple and clear. This result will be useful for the further work with multi variables of simple control formulation.

This paper proposes a new formulation which minimizes the differential entropy for an optimal control problem. The conventional criterion of the optimal regulator control is a standard quadratic cost function E[M{x(t)}2 + N{u(t)}2], where x(t) is a state variable, u(t) is an input value, and M and N are positive weights. However, increasing the number of the variables of the system it is complex to find the solution of the optimal regulator control. Therefore, the simplicity of the solution is required. In contrast to the optimal regulator control, we propose the minimum entropy control which minimizes a differential entropy of the weighted sum of x(t) and u(t). This solution is derived on the assumptions that the linear control and x(t)u(t) ≦ 0 are satisfied. As the result, the formula of the minimum entropy control is very simple and clear. This result will be useful for the further work with multi variables of simple control formulation.