Combining Modifications to Multinomial Naive Bayes for Text Classification

Abstract

Multinomial Naive Bayes (MNB) is a preferred classifier for many text classification tasks, due to simplicity and trivial scaling to large scale tasks. However, in terms of classification accuracy it has a performance gap to modern discriminative classifiers, due to strong data assumptions. This paper explores the optimized combination of popular modifications to generative models in the context of MNB text classification. In order to optimize the introduced classifier metaparameters, we explore direct search optimization using random search algorithms. We evaluate 7 basic modifications and 4 search algorithms across 5 publicly availably available datasets, and give comparisons to similarly optimized Multiclass Support Vector Machine (SVM) classifiers. The use of optimized modifications results in over 20% mean reduction in classification errors compared to baseline MNB models, reducing the gap between SVM and MNB mean performance by over 60%. Some of the individual modifications are shown to have substantial and significant effects, while differences between the random search algorithms are smaller and not statistically significant. The evaluated modifications are potentially applicable to many applications of generative text modeling, where similar performance gains can be achieved.