Tetsuji Kuboyama

Knowledge acquisition from graph structured data is an important task in machine learning and data mining. TTSP (Two-Terminal Series Parallel) graphs are used as data models for electric networks and scheduling. We propose an evolutionary learning method for obtaining characteristic multiple TTSP graph patterns with wildcards, from positive and negative TTSP graph data by clustering TTSP graphs.

Machine learning and data mining from tree structured data are studied intensively. We propose an evolutionary learning method for acquiring characteristic tag tree patterns with vertex labels and wildcards from positive and negative tree data, by using label information of positive examples. We report preliminary experimental results on our evolutionary learning method.

Machine learning from graph structured data are studied intensively. Many chemical compounds can be expressed by outerplanar graphs. The purpose of this paper is to propose a learning method for obtaining characteristic graph patterns from positive and negative outerplanar graph data. We propose a two-stage evolutionary learning method for acquiring characteristic multiple block preserving outerplanar graph patterns with wildcards from positive and negative outerplanar graph data, by using label information of positive examples. We report preliminary experimental results on our evolutionary learning method.

学習院大学のキャンパス内における無線 LAN の利用動向を分析した結果について報告する.

CWC (Combination of Weakest Components), a successful consistency-based feature selection algorithm, is known to outperform not only other consistency-based algorithms but many non-consistency-based filter algorithms in terms of time-efficiency and predictive accuracy. In this paper, we propose two methods to develop a new consistency-based feature selection algorithm based on CWC: One is to improve the predictive accuracy of CWC, and the other is to support the requirement of multiple outputs. With respect to the predictive accuracy, we found through intensive experiments that the order of investigating features to determine whether to eliminate the feature (CWC is a backward elimination algorithm) can affect the performance of CWC to a large extent, and identified the best order of the three orders that we investigated. On the other hand, when considering the way how we perform feature selection in the real world, we see that the role of feature selection is to provide candidates for the feature set that could explain the target problem the best, and we should examine their actual relevance separately in a manner specific to the problem. In this sense, it is desirable that a feature selection algorithm outputs multiple feature sets, but CWC can output only a single feature set per a dataset. Our idea to solve this problem is to run CWC multiple times taking advantage of its excellent time-efficiency, and we propose an algorithm for iterative execution of CWC using the outputs of the previous executions as feedback.