Some Research And Improvement On Associative Classification Algorithms

Since the associative classification was proposed, it has become a hot researchtopic and received amount of attentions from data mining area. The associativeclassification integrates the technology of association rule mining with classification,so it generates a fairly complete set of rules. A large number of studies andexperiments show that associative classification has a better classification result thantraditional classification algorithms, such as C4.5. Comparing to classificationalgorithms, such as neural network, SVM et al., associative classification is easier tounderstand in respect that its classifier is a set of rules. Moreover, it also has gooddata adaptability and robustness. Thus, associative classification has developed intoan important method of classification.However, there are also some problems difficult to overcome on associativeclassification algorithm, mainly due to the huge number of rules it generates. This notonly results in slowly computing and overloading system resource, but also that mostof the class associative rules have no contribution to classification result, sometimeseven weaken. This paper makes a relatively systemic study of associativeclassification algorithm, and presents the following work:(1) Propose An Associative Classification Method Based on Compact RulesPredictionSince a large number of class associative rule generated, classification algorithmneed to apply appropriate strategy to prune and select rules to build the classifier. Anypruning strategies used to establish guidelines for selecting rules and build theclassifier are based on the characteristics of the training data set. But thecharacteristics of prediction data may be not the same as training data, so the pruningmay go too far, leading the classifier over-fitting the training data set., as a result thatsome prediction data may enable to classified. Thus, the ability of classifier maybecome weaker. Therefore, This paper proposes a new associative classifier based oncompact rules prediction..the compact rules has a better probability to matches thedata than a single rule, strengthening the forecast function of a rule, so moreprediction data can be predicted by high rank rules.(2) Propose An Associative Classification algorithm Based on positive-and–negative-class information gainThe accuracy of Associative classification algorithm classification depends largely on how to select the classification rule sets. According that information gainmonotonically decreases to pattern’ support, DDPMine algorithm adopted FP-growth,and designed a branch-and-bound search for directly mining discriminative patternswhich can directly produce rules. Thus, DDPMine made revolutionary progress. Themeasurement of the rules should accurately reflect the information changes of the ruleconsequent. But for patterns, every class is equal, and the rule consequent is unknown,so the information gain of pattern can’t accurately represent the information changesof the rule consequent.. Thus, this paper proposes positive-and–negative-classinformation gain by treating rule consequent as positive class and other classes asnegative class, and theoretically proves its rationality. Base on this, we propose animproved algorithm Based on positive-and–negative-class information gain calledzFDDPMine...