Exploiting Textual Reviews for Recommendation Systems Improvement

Abstract

Nowadays, the development of the Web has motivated people to provide their opinions in textual form. Such textual information plays a core role in many methods of Recommendation Systems that can improve the accuracy of recommendations. On successful e-commerce sites, customers naturally tend to evaluate most products positively, which causes unbalanced distributions in the numbers of positive ratings compared to negative ratings. The result here is a decline in the accuracy of recommendation models when the ratios of these distributions converge. In this paper, we developed a recommendation model based on text reviews as additional information. In particular, five essential steps have shaped the proposed model. First, data preprocessing was applied, followed by two parallel steps, namely text classification, and topic modeling. Forth, a combination process was performed between the text polarity and the topic probabilities of the text, followed by the implementation of a process of text similarity. Finally, using the Naïve Bayes model after integrating the inferred adjustment weights in the classification score equation led to a recommendation of items most preferable to the target user. Extensive experiments were managed on three Amazon datasets, which showed that our proposed model exceeded all comparison methods in the Top-N recommendation task. Specifically, the improvement ratio ranged approximately from 10% to 61% on Musical Instruments, 26% on Automotive, and 22% on Amazon Instant Video in terms of F1-measure.