Aspect-based Rating Prediction with Deep Learning Based on the Combination of Aspect, Sentiment, and Semantic Features

Abstract

Online shopping is widespread, and online reviews have become essential for purchasing decisions; however, the massive amount of review data causes information overload for users. To solve this problem, scholars have used recommender systems to analyze users' purchase history and rating records and establish user preferences. In addition to rating information, text comments help conduct a precise examination of user preferences, thereby improving the prediction accuracy of recommender systems. Therefore, this study proposes the method of aspect-based rating prediction with deep learning. First, the aspect, sentiment, and semantic vectors of users and items are extracted from the review texts. Then, the deep learning method is used to extract and identify the user and item features. Finally, matrix factorization is applied to predict the items that users may be interested in, as well as their ratings. The results show that this method produced better predictions than those proposed by other relevant studies and improved recommendation accuracy.