Research On Fuzzy Association Rule Mining

In recent years, association-rule mining(ARM) has become an important issue since the derived rules can reveal the potential or implicit relationships among items in the binary databases. It takes two phases to find association rules. At first, it discovers the frequent itemsets based on the user-specific minimum support, and then it derives the association rules(ARs) from the discovered frequent itemsets based on the user-specific minimum confidence. This two-phase approach faces three major problems. First, the multiple database scans are required to generate the amounts of candidates in a level-wise way. Second, the traditional algorithms of ARM can only handle the binary databases. Third, it is a non-trivial task to set the appropriate minimum support or minimum confidence for discovering ARs. In the past, the type-1 fuzzy set was adopted to handle the quantitative database in which each item is represented by the linguistic terms with their corresponding fuzzy values based on the pre-defined membership functions. The type-2 fuzzy set can be concerned as the extension of type-1 fuzzy sets, which can be used to handle the uncertainly and reveal better inference than that of type-1 fuzzy sets.In the past, a level-wise algorithm was designed to mine the fuzzy frequent itemsets based on type-2 fuzzy sets, which requires more computations to generate-and-test candidate itemsets level-by-level. Besides, the multiple database scans are necessary to reveal the actual fuzzy frequent itemsets from the disc overed candidate itemsets. To solve the above problems, a fuzzy-list structure is presented to keep the necessary information for later mining process in this dissertation. A list-based fuzzy frequent itemset mining based on type-2 fuzzy sets(LFFMT2) algorithm is presented to mine fuzzy frequent itemsets based on the fuzzy-list structure. Two efficient pruning strategies are also developed to reduce the search space and speed up the computations for mining fuzzy frequent itemsets. From the conducted experiments, the designed LFFMT2 algorithm can reduce a great deal of computation in runtime. The number of traversal nodes for mining the fuzzy frequent itemsets can reduce by 23% in average, compared with the state-of-the-art Apriori-based algorithm. In further research, we improve the list-structure to make the pruning strategies being more efficient, and in the improved list-structure, we use the relative maximum value rather than the real maximum value, which leads to improve the efficiency of the prune strategies. In the experiments, compared with the former list, the number of traversal nodes is reduced by 14% in average with using the improved list structure, and thus, the runtime is also obviously reduced.Since it is a non-trivial task to set the appropriate minimum support threshold and minimum confidence threshold, a coherent relationship between items can thus be revealed based on the properties of propositional logic. In the second part of this dissertation, the type-2 fuzzy sets and the designed fuzzy-list structure used in the first part of this dissertation are both adopted to mine the fuzzy coherent association rules from the quantitative databases. Two pruning strategies are also developed to reduce the search space for exploring the fuzzy coherent association rules. From the experimental results, it can be observed that the designed algorithm has almost the same memory usage as the state-of-the-art algorithm but the runtime can be greately reduced and the number of traversal nodes can be reduced b y at least 48% in average.