Naive Bayes classifiers estimate posterior probabilities poorly (Zhang, 2004). In this paper, we propose a modification to the Naive Bayes classification algorithm which improves the classifier's posterior probability estimates without affecting its performance. Since the modification involves the use of the reciprocal of the perplexity of the class-conditional feature probabilities, we call the resulting classifier the Perplexed Bayes classifier. We demonstrate that the modification results in better calibrated posterior probabilities on a gender categorization task.
展开▼
