An approach to improve the accuracy of rating prediction for recommender systems

Abstract

Sentiment analysis is critical for classifying users on social media and reviewing products through comments and reviews. At the same time, rating prediction is a popular and valuable topic in research on recommendation systems. This study improves the accuracy of ratings in recommendation systems through the combination of rating prediction and sentiment analysis from customer reviews. New ratings have been generated based on original ratings and sentiment analysis. Experimental results show that in almost all cases, revised ratings using a deep learning-based algorithm called LightGCN on 7 various real-life datasets improve rating prediction. In particular, rating prediction metrics (RMSE and MAE, R2, and explained variance) of the proposed approach (with revised ratings) are better than those of the typical approach (with unrevised ratings). Furthermore, evaluating ranking metrics (also top-k item recommendation metrics) for this model also shows that our proposed approach (with revised ratings) is much more effective than the original approach (with unrevised ratings). Our significant contribution to this research is to propose a better rating prediction model that uses a supplement factor sentiment score to enhance the accuracy of rating prediction.