Decision Rule Clustering—Comparison of the Algorithms

Abstract

In this paper, we present the complexity of decision rule clustering. When the rules are first clustered, then in the inference process we do review only the representatives of the rule clusters. This shortens the inference time significantly, because we search only k rule cluster representatives instead of n rules, where $k << n$. The main goal of the research was to examine the two well-known clustering algorithms: the K-means and the AHC, in the context of rule-based knowledge representation. We tested different clustering approaches, distance measures, clustering methods, and values for the parameter representing the number of created rule clusters. We studied the clustering time and cluster quality indices. This paper is the first step of a more extensive study. After we have checked which algorithm clustering the rules faster in the knowledge base, we will propose our own version of the inference algorithm for rule clusters, a modification of the classic forward chaining process (on rules). Next, we will carry out experiments that are a continuation of those carried out for this work. These experiments will focus on analyzing the times of the classical inference process and its modification and the efficiency of inference, which will be measured, among others, by the frequency of successful conclusions of inference for both versions of inference algorithms. In this way, we will check whether, by clustering the rules and generating the conclusions on clusters of rules while significantly reducing the reasoning time, we can maintain high efficiency of reasoning.