Review text based rating prediction approaches: preference knowledge learning, representation and utilization

Numerous rating prediction approaches have exploited users’ review texts to learn the associated preference knowledge or content semantics in order to make more accurate predictions. Such approaches either involve traditional machine learning techniques or deep learning practices to learn, extract and represent different preference knowledge such as review topics, review sentiments, linguistic aspects and feature words. With the huge number of users’ review texts on products or services and as researchers propose new rating prediction methods which utilize different preference knowledge, it’s necessary to review the methods, the acquired knowledge and how such methods make predictions. This study unveils comprehensive overview of the acquired preference knowledge and how the rating prediction approaches learn, represent and utilize such knowledge. Associated prediction methods were analyzed and presented along two perspectives: traditional machine learning; and deep learning practices. This paper not only evaluates the influence of the acquired preference knowledge in extending or regulating base methods but also identifies associated challenges in predicting ratings. Selected publications were analyzed to reveal different tactics which rating prediction approaches utilize to resolve data sparsity along with cold start problems. Finally, a discussion about possible future trends is presented. The study suggests that application of effective techniques for learning, representing and utilizing preference knowledge can improve prediction accuracy of the models. It also advocates that different combinations of the acquired preference knowledge can enhance prediction performance of the rating prediction approaches.