Incremental induction rules clustering

Abstract

The current world wide web is featured by a huge volume of knowledge, making it possible to apply knowledge mining to extract meta-knowledge. This paper explores this possibility and considers knowledge discovery process acceleration. Given that knowledge is extracted from data, knowledge mining process would be similar to data mining. However, knowledge representation is more complex than data representation. Homogeneous knowledge, such as induction rules, should thus be mined first. An extension of k-means algorithm is proposed, which clusters induction rules using a new similarity measure. On the other hand, induction rules are continually and dynamically acquired, e.g. agent concept. It is more efficient for the discovery process to incrementally mine induction rules. Three incremental induction rule clustering approaches are developed. These approaches have been tested using three benchmarks, and their clustering performance has been investigated. Results are satisfactory and show from 70% to 90% of success rate.