Fuzzy Classification Based On Fuzzy Association Rule Mining

Abstract

In fuzzy classification of high-dimensional datasets, the number of fuzzy rules increases exponentially with the increase of attributes. Fuzzy association rule mining with appropriate threshold values can help to design a fuzzy classifier by significantly decreasing the number of interesting rules. In this dissertation, we investigate the way to integrate fuzzy association rule mining and fuzzy classification. First, the framework of fuzzy association rule mining is presented which incorporates fuzzy set modeling in an association rule mining technique. It avoids the sharp boundary problem caused by arbitrary determination of intervals on the domain of quantitative attributes, meanwhile presenting natural and clear knowledge in the form of linguistic rules. We study the impact of different fuzzy aggregation operators on the rule mining result. The selection of the operator should depend on the application context. Based on the framework of fuzzy association rule mining, we propose a heuristic method to construct the fuzzy classifier based on the set of fuzzy class association rules. We call this method the FCBA approach, where FCBA stands for fuzzy classification based on association. The objective is to build a classifier with strong classification ability.

In the FCBA approach, we use the composite criteria of fuzzy support and fuzzy confidence as the rule weight to indicate the significance of the rule. Through our study, we find it is important to find a good combination of these two rule interestingness threshold values. The classification of each record is achieved by applying the classic fuzzy reasoning method, in which each record is classified as the consequent of the rule with the maximum product of the compatibility grade and the rule weight. We use the well-known classification problems such as the Iris dataset, and high-dimensional classification problem, the Wine dataset to compare the proposed FCBA approach with other non-fuzzy and fuzzy classification approaches. The empirical study shows that the FCBA approach performs well on these datasets on both accuracy and interpretability.