Shigeaki Sakurai

An e-mail analysis method based on text mining techniques

This paper proposes a method employing text mining techniques to analyze e-mails collected at a customer center. The method uses two kinds of domain-dependent knowledge. One is a key concept dictionary manually provided by human experts. The other is a concept relation dictionary automatically acquired by a fuzzy inductive learning algorithm. The method inputs the subject and the body of an e-mail and decides a text class for the e-mail. Also, the method extracts key concepts from e-mails and presents their statistical information. This paper applies the method to three kinds of analysis tasks: a product analysis task, a contents analysis task, and an address analysis task. The results of numerical experiments indicate that acquired concept relation dictionaries correspond to the intuition of operators in the customer center and give highly precise ratios in the classification.

Analysis of daily business reports based on sequential text mining method

This paper proposes a new method that discovers characteristic sequential patterns in textual data. The data are composed of three kinds of information: time information, attributes, and text. The method gathers items of the data with the same attribute values, arranges the gathered items in order of the time, and generates sequences. The method also extracts events from each text by using a text mining method. Finally, the method discovers characteristic sequential patterns, composed of sets of events, from sequences by a sequential mining method. In this paper, we apply the method to business reports collected by our sales force automation system and try to discover characteristic sequential patterns. We verify whether the patterns are valid by investigating texts relating to the patterns.

Discovery of Sequential Patterns Coinciding with Analysts’ Interests

This paper proposes a new sequential pattern mining method. The method introduces a new evaluation criterion satisfying the Apriori property. The criterion is calculated by the frequency of the sequential pattern and the minimum frequency of items included in the items. It extracts sequential patterns that can be rules predicting future items with high probability. Also, the method introduces new constraints. The constraints extract item sets composed of items whose attributes are different and extracts sequential patterns composed of item sets whose attribute sets are equal to one another. The proposed method efficiently discovers sequential patterns coinciding with analysts’ interests by combining the criterion and the constraints. The paper verifies the effectiveness of the proposed method by applying it to medical examination data.

A New Approach For Discovering Top-K Sequential Patterns Based On The Variety Of Items

Abstract This paper proposes a method that discovers various sequential patterns from sequential data. The sequential data is a set of sequences. Each sequence is a row of item sets. Many previous methods discover frequent sequential patterns from the data. However, the patterns tend to be similar to each other because they are composed of limited items. The patterns do not always correspond to the interests of analysts. Therefore, this paper tackles on the issue discovering various sequential patterns. The proposed method decides redundant sequential patterns by evaluating the variety of items and deletes them based on three kinds of delete processes. It can discover various sequential patterns within the upper bound for the number of sequential patterns given by the analysts. This paper applies the method to the synthetic sequential data which is characterized by number of items, their kind, and length of sequence. The effect of the method is verified through numerical experiments.

Rule discovery from textual data based on key phrase patterns

This paper proposes a new method for discovering rules from textual data. The method decomposes textual data into word sets by using lexical analysis, generates training examples from both key phrase relations extracted from the word sets by using key phrase patterns and text classes given by the user, and acquires key phrase relation rules from the examples by using a fuzzy inductive learning algorithm. The method is also able to deal with textual data that requires word segmentation, such as Japanese text. This paper reports on the application of the method to e-mail analysis tasks for a customer center. The e-mails are written in Japanese and have two analytical criteria: a product criterion and a contents criterion. We evaluate the acquired rules in each criterion.

Inductive Learning of a Knowledge Dictionary for a Text Mining System

A text mining system using domain-dependent dictionaries efficiently analyzes text data. The dictionaries store not only important words for the domains, but also rules composed of some important words. The paper proposes a method that automatically acquires the rules from the text data and their classes by using a fuzzy inductive learning method. Also, in order to infer a class corresponding to new text data, the paper proposes an inference method based on the acquired fuzzy decision tree. Moreover, the efficiency of the methods is verified through numerical experiments using more than 1,000 daily business reports concerning retailing.

A Discovery Method of Trend Rules from Complex Sequential Data

This paper proposes a method that discovers trend rules from complex sequential data. The rules represent relationships among evaluation objects, keywords, and changes of numerical values related to the evaluation objects. The data is composed of numerical sequential data and text sequential data. The method extracts frequent patterns from transaction sets based on the changes. Also, it regards combinations of the patterns and the changes as trend rules. This paper applies the method to data sets composed of stock data and news headlines. Lastly, this paper compares the method with a method based on the random selection and shows the effect of the proposed method.

A clustering method of bloggers based on social annotations

This paper proposes a method that divides bloggers to clusters according to their interests. The method calculates similarities between the bloggers based on three steps. That is, the method calculates similarities between target objects discussed in blog articles based on social annotations. It calculates similarities between impressions related to the target objects based on impression words included in blog articles. Here, products, works and services are examples of the target objects. Lastly, the method calculates similarities between the bloggers by combining the results of the methods first and second calculation steps, and divides the bloggers to clusters based on the similarities. The paper applies the method to the Commutents data and the Yahoo! Japan Movie data, and verifies the effectiveness of the method.

Discovery of Association Rules from Data including Missing Values

This paper proposes a method that deals with missing values in the discovery of association rules. The method deals with items composed of attributes and attribute values. The method calculates two kinds of support. One is characteristic support and the other is possible support. The former is based on the number of examples that do not include missing values in attributes composing target items. The latter is based on the number of examples that do not include missing values in all attributes. The method extracts all item sets whose characteristic supports are larger than or equal to the predefined threshold. The paper evaluates the proposed method by comparing it with the previous method and verifies the effect of the proposed method.

Sequential pattern mining based on a new criteria and attribute constraints

This paper proposes the sequential interestingness as a new evaluation criterion that evaluates a sequential pattern corresponding to the interests of analysts. The sequential pattern is composed of rows of item sets. The criterion satisfies the apriori property. Also, this paper proposes three attribute constraints. These constraints can naturally evaluate relationships of attributes both in an item set and between continuous item sets. In addition, this paper proposes a mining method incorporating the criterion and the constraints. The method can efficiently discover all sequential patterns whose sequential interestingness is larger than or equal to a threshold and that satisfy the constraints. Lastly, this paper verifies the effectiveness of the proposed method by applying the method to medical examination data.