An experimental study on the performance of collaborative filtering based on user reviews for large-scale datasets

Collaborative filtering (CF) approaches generate user recommendations based on user similarities. These similarities are calculated based on the overall (explicit) user ratings. However, in some domains, such ratings may be sparse or unavailable. User reviews can play a significant role in such cases, as implicit ratings can be derived from the reviews using sentiment analysis, a natural language processing technique. However, most current studies calculate the implicit ratings by simply aggregating the scores of all sentiment words appearing in reviews and, thus, ignoring the elements of sentiment degrees and aspects of user reviews. This study addresses this issue by calculating the implicit rating differently, leveraging the rich information in user reviews by using both sentiment words and aspect-sentiment word pairs to enhance the CF performance. It proposes four methods to calculate the implicit ratings on large-scale datasets: the first considers the degree of sentiment words, while the second exploits the aspects by extracting aspect-sentiment word pairs to calculate the implicit ratings. The remaining two methods combine explicit ratings with the implicit ratings generated by the first two methods. The generated ratings are then incorporated into different CF rating prediction algorithms to evaluate their effectiveness in enhancing the CF performance. Evaluative experiments of the proposed methods are conducted on two large-scale datasets: Amazon and Yelp. Results of the experiments show that the proposed ratings improved the accuracy of CF rating prediction algorithms and outperformed the explicit ratings in terms of three predictive accuracy metrics.

Automated summary

Collaborative filtering (CF) approaches generate user recommendations based on user similarities, but sparse or unavailable user ratings can be supplemented with implicit ratings derived from user reviews using sentiment analysis. This study proposes four methods to calculate implicit ratings by incorporating sentiment degrees and aspect-sentiment word pairs, and evaluates their effectiveness in enhancing CF performance.