Bayesian Classifier Modeling for Dirty Data

摘要

Bayesian classifiers have been proven effective in many practical applications. To train a Bayesian classifier, important parameters such as prior and class conditional probabilities need to be learned from datasets. In practice, datasets are prone to errors due to dirty (missing, erroneous or duplicated) values, which will severely affect the model accuracy if no data cleaning task is enforced. However, cleaning the whole dataset is prohibitively laborious and thus infeasible for even medium-sized datasets. To this end, we propose to induce Bayes models by cleaning only small samples of the dataset. We derive confidence intervals as a function of sample size after data cleaning. In this way, the posterior probability is guaranteed to fall into the estimated confidence intervals with constant probability. Then, we design two strategies to compare the posterior probability intervals if overlap exists. Extension to semi-naive Bayes method is also addressed. Experimental results suggest that cleaning only a small number of samples can train satisfactory Bayesian models, offering significant, improvement in cost over cleaning all of the data and significant improvement on precision, recall and F-Measure over cleaning none of the data.