Classification with Maximum Entropy Modeling of Predictive Association Rules

Abstract

This paper presents a new classification model in which a classifier is built upon predictive association rules (PARs) and the maximum entropy principle (maxent). In this model, PARs can be seen as confident statistical patterns discovered from training data with strong dependencies and correlations among data items. Maxent, on the other hand, is an approach to build an estimated distribution having maximum entropy while obeying a potentially large number of useful features observed in empirical data. The underlying idea of our model is that PARs have suitable characteristics to serve as features for maxent. As a result, our classifier can take advantage of both the useful correlation and confidence of PARs as well as the strong statistical modeling capability of maxent. The experimental results show that our model can achieve significantly higher accuracy in comparison with the previous methods.