Weakening Feature Independence of Naïve Bayes Using Feature Weighting and Selection on Imbalanced Customer Review Data

E-commerce sites have provided review section for users to take advantage of and express their opinion about products or services. Decision makers, on other hand, can also take advantage of the abundant reviews to analyze which aspects of products or services to be improved, which is known as sentiment analysis. naïve Bayes (NB) is popular method for sentiment analysis because it is considerably faster than other methods but has comparable performance. One weakness of NB however, is that NB assumes each feature is independent with other features. This assumption is not fulfilled in sentiment analysis because terms are correlated with others. Two approaches, i.e. feature weighting (FW) and feature selection (FS) are used to weaken this assumption. Two approaches use genetic algorithm (GA) to find optimal weights and subset based on correlation and odds ratio to take imbalanced review data into account. Experiments on Women Ecommerce Clothing Review dataset show that FW approach has comparable results to non-weighted NB and FS yield worse results than NB. It can be concluded that proposed FW and FS scheme cannot improve standard NB.