Research On Emerging Pattern Classification Based On Lazy Learning

As the rule form of class’s private features, Emerging patterns are more powerful in classification than normal rules. However, large-scale EP set includes a lot of redundant rules and it is difficult to design a strategy that choosing EPs effectively. According to Minimal Description Length principle (MDL), traditional EP-based classification method is to reserve the most expressive short-rules to construct classifier which own small number of EP. Nonetheless, these methods still have some problems, such as EP set being simplified overly and EP being chose unreasonably. For overcoming problems above, lazy learning strategy is proposed to transform traditional EP-based classifier.(1) LLEP classifier based on lazy learning strategy, has been proposed to solute some problems in traditional EP-based classification methods, such as EP set being simplified overly and EP being chose unreasonably. New growth rate and computational formula are proposed in LLEP method to discover EP accurately in imbalance data sets. Lazy learning strategy has been imported to postpone choosing EP and constructing classifier to classification stage. In order to find powerful EP for classification, long EPs have been given priority to construct classifier. Meanwhile, the coverage of classifier has been considered. For improving LLEP performance further, compressing method based on equal class has been proposed in this thesis, which can be used to exclude those redundant EP patterns. The experiments show that, LLEP based on lazy learning strategy is more accurate than traditional EP-based classifier.(2) The causal relevance is an important part of association relationship, and it can describe essential relationship between attributes. Causal emerging pattern has been imported to overcome limit that Markov Blanket can only describe causal relationship between single attribute, because it can describe causal relationship between attribute sets and class labels. Chi-square statistic has been used to calculate association level between attribute sets and class labels, and EPs with high association level would be regarded as causal emerging patterns. On the basis of LLEP method, CEP classifier based on causal emerging pattern would be constructed. The experiments show that, CEP classifier is more accurate than twelve traditional classifiers, and CEP classifier also uses less EP patterns than other EP-based classifiers.