Fuzzy association rule mining and classifier with chi-squared correlation measure using genetic network programming
Karla Taboada, Shingo Mabu, Eloy Gonzales, Kaoru Shimada, Kotaro Hirasawa · 2009 ICCAS-SICE · 2009
One of the most important issues in any association rule mining is the interpretation and evaluation of discovered rules. Thus, most algorithms employ the support-confidence framework for evaluating association and classification rules. Unfortunately, recent studies show that the support and confidence measures are insufficient for filtering out uninteresting association rules, for instance, even strong association rules can be uninteresting and misleading. To deal with this limitation, the support-confidence framework can be suplemented with additional interestingness measures based on statistical significance and correlation analysis. In this paper, a novel fuzzy association rule-based classification approach is proposed, where χ f2 is applied as a correlation measure. The algorithm is based on Genetic Network Programming (GNP) and discover comprehensible fuzzy association rules potentially useful for classification. GNP is an evolutionary optimization algorithm that uses directed graph structures as genes instead of strings and trees of Genetic Algorithms (GA) and Genetic Programming (GP), respectively. This feature contributes to creating quite compact programs and implicitly memorizing past action sequences. The proposed model consists of two major phases: 1) generating fuzzy class association rules by using GNP, 2) building a classifier based on the extracted fuzzy rules. In the first phase, χ f2 is used for computing the correlation of the rules to be integrated into the classifier. In the second phase, the χ f2 value is used as a weight of the rule when calculating the matching degree of the rule with new data. The performance of the proposed algorithm has been compared with other relevant algorithms and the experimental results have shown the advantages and effectiveness of the proposed model.