Using Clustering Algorithms to Improve the Production of Symbolic-Neural Rule Bases from Empirical Data

Abstract

Neurules are a kind of integrated rules integrating neurocomputing (via the adaline unit) and production rules. A neurule base is modular and natural, in contrast to existing connectionist knowledge bases, a comparable type of integrated knowledge bases. In producing neurules from an empirical data training set, the inability of the adaline unit to classify non-separable training data should be faced. The general approach followed is consecutively splitting the training set into two subsets, according to a splitting strategy, until (sub)sets of separable data are produced; then as many neurules as the resulted subsets are produced. In this paper, we present and experimentally evaluate six splitting strategies applied to the production process of a neurule base, three of which are based on clustering algorithms suitable for categorical data (i.e., 2-medoids, 2-modes and COOLCAT). Experiments were performed using 18 different distance or similarity metrics suitable for categorical data. No such an extensive comparison of distance/similarity metrics has been made so far. The strategy based on 2-modes generally performs better than the other strategies by applying alternative cluster center initialization methods. Specific distance/similarity metrics also provide better results.