AbstractsComputer Science

Association rule mining is commonly utilized to extract useful information from given data. Since items are usually with quantities in real-world transaction databases, the fuzzy set theory is applied to many mining approaches for deriving fuzzy association rules. In the past, fuzzy mining mainly focused on type-1 membership functions. In this thesis, we attempt to use type-2 membership functions for mining. Type-2 fuzzy sets are generalization of type-1 fuzzy sets and are able to handle more uncertainty than type-1. An interval type-2 fuzzy association rule mining approach is first proposed in this thesis. Rules are mined by predefined interval type-2 membership functions. The quantitative transactions are transformed into fuzzy values according to the corresponding type-2 membership functions. The interval type-2 fuzzy values will be reduced to type-1 values by a centroid type reduction method in order to induce fuzzy association rules. Since membership functions are usually assumed to be known in advance in most of the fuzzy data mining approaches, thus a GA-based type-2 fuzzy association rule mining is proposed to learn appropriate type-2 membership functions. The type-2 membership functions of each item are encoded as a chromosome and appropriate genetic operators are designed to find good solutions. In order to further enhance the quality of mining results, another GA-based representation, the 2-tuple linguistic representation, is also proposed. It adopts a different tuning mechanism and a modified evaluation function for the chromosomes to evolve. Experiments are also made to show the effectiveness of the proposed approaches. From the experimental results, the proposed approaches can mine more rules than using type-1 membership functions, and the qualities of rules are improved as well. Advisors/Committee Members: Wen-yang Lin (chair), Tzung-Pei Hong (committee member), Ming-chao Chiang (chair), Chun-Hao Chen (chair).