A decomposition of classes via clustering to explain and improve Naive Bayes

摘要: We propose a method to improve the probability estimates made by Naive Bayes avoid effects of poor class conditional probabilities based on product distributions when each spreads into multiple regions. Our approach is applying clustering algorithm subset examples that belong same class, and consider cluster as its own. Experiments 26 real-world datasets show significant improvement in performance decomposition process applied, particularly mean number clusters per large.