Knowledge management system: Combination of experts' knowledge and automatic improvement

Abstract

This paper proposes a knowledge management system that acquires knowledge from time series data by using users background knowledge. To obtain if-then rules as knowledge, we apply decision tree learning. However usual methods of decision tree learning targets discrete value data, thus a new approach is needed for dealing with this type of data. Experts forecast future events by using their knowledge. They, in typical cases, focus on a set of useful patterns and then apply knowledge relevant to them. We apply this idea into the framework of decision tree learning. We prepare a set of patterns, which is called clues, and then express time series data in terms of the clues. Thus the clues are attributes by which features of data are described. In addition, to make a better prediction with the learning process, we develop a mechanism that improves the quality of the clues. The essential idea of the mechanism is based on a genetic algorithm. The clue is evaluated by using entropy of information theory, and is improved by GA operators. We can obtain new knowledge from improved clues and the extracted decision tree. This paper details the system and results of the experiment. © 2013 Springer-Verlag GmbH.