Hierarchical Scheme for Assigning Components in Multinomial Naive Bayes Text Classifier
Nghia T. Nguyen, Kôichi Yamada, Izumi Suzuki, Muneyuki Unehara · 2018
Naive Bayes Classifier is an effective and easy method in general, and is used popularly in document classification. It's also a particularly useful classifier for large dataset. Naive Bayes assigns instances to a class label through maximizing the posterior probability assuming that the instance features are independent of each other. Using this model, many-to-one assumption was proposed to deal with the case where a class may be constituted of many different components. An issue of this approach is how to construct these components on each class. In this paper, we propose a simple heuristic solution of applying a hierarchical tree for assigning components to classes. Through evaluation experiments, we find that the proposed approach shows better performance than the conventional one when the training data size is large enough. Some other issues in practicing many-to-one assumption are also discussed.