Abstract
The naive Bayes (NB) learning algorithm is widely applied in many fields, particularly in text classification. However, its performance decreases when it is used in domains where its naive assumption is violated or when the training set is too small to find accurate estimations of the probabilities. In this study, we propose a lazy fine-tuning naive Bayes (LFTNB) method to address both problems. We propose a local fine-tuning algorithm that uses the nearest neighbors of a query instance to fine-tune the probability terms used by NB. Applying the nearest neighbors only makes the independence assumption more likely to be valid, whereas the fine-tuning algorithm is used to find more accurate estimations of the probability terms. The performance of the LFTNB approach was evaluated using 47 UCI datasets. The results show that the LFTNB method achieves superior performance than classical NB, eager FTNB, and k-nearest neighbor algorithms. We also propose eager and lazy fine-tuning versions of powerful NB-based text classification algorithms, namely, multinomial NB, complement NB, and one-versus-all NB. The empirical results using 18 UCI text classification datasets show that the proposed methods outperform untuned versions of these algorithms. (C) 2020 Elsevier B.V. All rights reserved.