EARC: Evidential Association Rule-based Classification

Abstract As an extension of classical fuzzy rule-based classification, the belief rule-based classification is a promising technique for handling hybrid information with multiple uncertainties in real-world applications. However, the antecedent structure of each resultant rule is fixed and hence may cause overfitting in small instance cases, while some resultant rules are also redundant due to the similarity of neighboring rules. Here, an evidential association rule-based classification method, called EARC, is developed by integrating evidential association rule mining and classification to obtain an accurate and compact classification model. First, new measures of evidential support and confidence are proposed to represent rule interestingness. Then, a three-stage rule mining algorithm is developed to generate a set of evidential classification association rules, including Apriori-based frequent fuzzy itemsets searching for discovering all possible antecedents, evidential consequents deriving in the belief function framework, and reliable rule extracting with measures of evidential support and confidence. Further, to make the classification efficient, the procedures of rule prescreening and rule selection are presented for deleting redundant rules and obtaining an accurate classifier, respectively. At last, an improved belief reasoning process is presented for classifying each input instance by combining the top K activated rules. Experimental results based on real-world datasets demonstrate the superiority of the proposed method on classification accuracy and interpretability.