Recently, there are many opportunities to acquire text information as the quantity of electronical information increases. Data classification or clustering methods are widely adapted in order to acquire various information effectively from an enormous dynamic text data set like Web pages. However, we cannot see various information because ordinal clustering methods connect texts and many texts are concentrated into a single cluster. In this study, we propose a recursive clustering method to avoid such bias by integrating a set of texts, included in a cluster, into a single text. An interface that we can comprehend a result of clustering intuitively and can explore information is required to grasp an overview of data and to be led to a new idea. According to the experimental results, the proposed method could construct clusters that are not biased. Test subjects could find information widely by using a map visualizing clustering results.

View full abstract

These days, online communication tools, such as e-mail, bulletin board systems, and social network services, are frequently used in our daily lives. Especially, bulletin board systems are used not only as communication tools but also as knowledge database that includes conclusions of discussion. herefore, opinion exchange and information retrievals should be supported so that we can take part in a communication comfortably and see effective information within a good mood. However, some bulletin board systems may be in a bad mood because of anonymous comments that hurt other user's mind. In this paper, a system that identifies communication moods in bulletin board systems is proposed. The system evaluates moods in sequential comments related to each topic and outputs those to users. We can easily find communication in good moods or in bad moods by seeing output of the system.

View full abstract

In this paper, we analyze web access logs of a web site that synchronizes with free magazine for working woman in Tokyo, in order to investigate browse behavior of these webpages. In general, it is difficult to identify gender and age of access users using web access logs. We try to extract browse behavior of specific users using web access logs that synchronizes with free magazine. As analysis results, we discover characteristic behaviors of users.

View full abstract

In the practice of data mining, it sometimes happens that the user cannot get business meanings only from the extracted rules and cannot derive insights. For such situation, in this research, we use text mining for the trend analysis and get results that can be utilized for filtering data mining results. In order to acquire valuable knowledge through text mining, it is generally critical to identify important expressions to be monitored and compared within the textual data. In this paper, we propose the method to reach to the trends leading to the insights in the marketing analysis and show how we can utilize the text mining results for filtering data mining results. Through the analysis of real textual data for the market analysis of a garbage disposal, we will show that we can effectively get insights through the proposed method.

View full abstract

In data mining approach, the quantitative attributes should be appropriately dealt with as well as the Boolean attributes corresponding to various applications and target areas. This paper presents a fast algorithm for extracting fuzzy association rules from a massive database. The objective of the algorithm is to improve the computational time of mining for actual applications. In this paper, we define redundancy of fuzzy association rules as a new concept for mining and essential theorems concerning with the redundancy of fuzzy association rules. Then, we propose a basic algorithm based on the Apriori algorithm for rule extraction utilizing output attribute specifications and redundancy of the extracted rules. The performance of the algorithm is evaluated through numerical experiments using benchmark data. From the results, the proposed method is found to be promising in terms of computational time and redun dant rule pruning.

View full abstract

A recommendation system enables us to take information from huge datasets about tastes effectively. Many cryptographical protocols for computing privacy-preserving recommendation without leaking the privacy of users are proposed. However, the current issue is the large computational overhead depending the number of users and hence, the application of the protocol is limited within small communities. In this paper, we propose some efficient schemes reducing the preference matrix of the sets of items and users. In the proposed schemes, users get their ratings encrypted by a public key of trusted authorities and submit to a public server so that the ciphertext of ratings are available to any users. The user wishing to have the recommendation performs some precomputations of these using the homomorphic property of public key algorithm and then sends the resulting ciphertext to the set of trusted servers. Having sampled some users rating in secure way, it divides the set of users into smaller groups which results in reduction of matrix in low dimension.

View full abstract

This paper discuss bipedal walking control based on human skills. Concretely, we focus on walking control described in words, for instance, “we shift our contact point with the ground further forward when we walk faster.” or “We accelerate the body by extending the kenee joint of supporting leg after the body's COG (center of gravity) passes over the supporting leg.” Replacing these rules by fuzzy rules, a model for bipedal walking control is proposed. In this paper, I analyze walking control by a inverted pendulum model. Using this model, well defined rules on extend force of the supporting leg, the landing point of the swing leg, and conditions for leg-support-exchange allow stable walking. The proposed system enables walking velocity control, walking direction control and height control of COG by using 13 rule tables concerned with above-mentioned conditions. Through computer simulations using ODE (Open Dynamics Engine), validity of proposed model is confirmed.

View full abstract

Target enclosure behavior of autonomous robots is useful for many practical applications, for example, surveillance of disaster sites. Scalability is one of important capabilities for autonomous robots because a larger group has more robust for breakdown, accidents and failure. However, it is more difficult to operate a larger group of robots because it needs higher capabilities to recognize teammates individually. In this paper, to achieve high scalability of target enclosure model, we show a new condition for Takayama's enclosure model. Originally, their model needs a static relation between agents. However, the robots can form an enclosure even dynamical topology based on nearest neighbor graph, under which robots do not need the recognition capabilities. The authors show this issue by analytic discussion based on switched systems and a series of computer simulations.

View full abstract

This paper proposes a new optimization algorithm based on Artificial Bee Colony (ABC) algorithm for large-scale engineering design problems, which can detect plural acceptable solutions. The proposed algorithm is evaluated through the numerical experiments on benchmark functions with many local minima, such as Rastrigin function, Schwefel function and Griewank function, and its development potential is discussed. In the numerical experiments, the performances of the proposed algorithm are compared with those of the existing optimization ones detecting multiple solutions to verify its effectiveness.

View full abstract

In this paper, we develop a Bayesian network model that infers the reproducibility of preferences for pairs of fashion model stimuli from eye movements. Three types of fashion model pairs (i. e. , high score vs. high score, low score vs. low score, and high score vs. low score) are provided according to the evaluation score of each individual model. Using the Bayesian network model, we examine the relation between reproducibility and eye movements for each type of pair of fashion model stimuli. In the case of large decision latency, the probabilities of reproducibility differ considerably in two gazing patterns, i. e. , gazing frequently at the stimulus that is chosen or at the stimulus that is not chosen. In order to determine the cause of this difference, the occurrence probabilities of uncertain feelings in preference decisions are inferred from eye movements. It turns out that the two gazing patterns correspond to distinct attitudes toward preference decisions.

View full abstract

There are numerous onomatopoeic words unique to the Japanese language. Those words are indispensable to make smoother communication in Japanese. Though native speakers of Japanese normally use onomatopoeias without thinking, they can hardly restate them in different terms, let alone explain what they mean. It is said that onomatopoeias are quite hard to master for people studying Japanese. One way to do this is the exposure to many sentences that contain onomatopoeia, and hence we propose an automatic onomatopoeia sentences extraction method based on an analysis of example sentences. As a result, our method extracts 15,437 onomatopoeia sentences from 41,315 blog entries and achieves an accuracy of 96.2% (481/500).

View full abstract

In this paper, the FCMdd-based linear fuzzy clustering method is applied to non-Euclidean relational data in conjunction with data transformation like NERF c-Means with the goal of constructing local feature map in each local area. Two experimental results are shown to demonstrate how the local feature map can be constructed considering local structures of non-Euclidean relational data.

View full abstract