Bayesian Classifier Based on Frequent Item Sets Mining

Nave Bayesian classifier provides a simple and effective way to classifier learning, but its assumption on attribute independence is often violated in real-world applications. To alleviate this assumption and improve the generalization ability of Nave Bayesian classifier, many works have been done cy researchers. AODE ensembles some one-dependence Bayesian classifiers and LB selects and combines long item sets providing new evidence to compute the class probability. Both of them achieve good performance, but higher order dependence relations may contain useful information for classification and limiting the number of item sets used in classifier may restricts the benefit of item sets. For this consideration, a frequent item sets mining-based Bayesian classifier, FISC (frequent item sets classifier), is proposed. At the training stage, FISC finds all the frequent item sets satisfying the minimum support threshold min_sup and computes all the probabilities that may be used at the classification time. At the test stage, FISC constructs a classifier for each frequent item set contained in the test instance, and then classifies the instance by ensembling all these classifiers. Experiments validate the effectiveness of FISC and show how the performance of FISC varies with different min_sup. Based on the experiment result, an experiential selection for min_sup is suggested.