Review-based collaborative recommender system using deep learning methods

Abstract

Recommender systems have been widely adopted to assist users in purchasing and increasing sales. Collaborative filtering techniques have been identified to be the most popular methods used for the recommendation system. One major drawback of these approaches is the data sparsity problem, which generally leads to low performances of the recommender systems. Recent development has shown that user review texts can be exploited to tackle the issue of data sparsity thereby improving the accuracy of the recommender systems. However, the problem with existing methods for the review-based recommender system is the use of handcrafted features which makes the system less accurate. Thus, to address the above issue, this study proposed collaborative recommender system models that utilize user textual reviews based on deep learning methods for improving predictive performances of recommender systems. To extract the product aspects to mine users‟ opinion, an aspect extraction method was first developed using a Multi-Channel Convolutional Neural Network. An aspect-based recommender system was then designed by integrating the opinions of users based on the product aspects into the collaborative filtering method for the recommendation process. To further improve the predictive performance, the fine-grained user-item interaction based on the aspect-based collaborative method was studied and a sentiment-aware recommender system was also designed using a deep learning method. Extensive series of experiments were conducted on real-world datasets from the Semeval-014, Amazon, and Yelp reviews to evaluate the performances of the proposed models from both the aspect extraction and rating prediction. Experimental results showed that the proposed aspect extraction model performed better than compared methods such as rule-based and the neural network-based approaches, with average gains of 5.2%, 12.0%, and 7.5% in terms of Precision, Recall, and F1 score, respectively. Meanwhile, the proposed aspect-based collaborative methods demonstrated better performances compared to benchmark approaches such as topic modelling techniques with an average improvement of 6.5% and 8.0% in terms of the Root Means Squared Error (RMSE) and Mean Absolute Error (MAE), respectively. Statistical T-test was conducted and the results showed that all the performance improvements were significant at P<0.05. This result indicates the effectiveness of utilizing the multi-channel convolutional neural network for better extraction accuracy. The findings also demonstrate the advantage of utilizing user textual reviews and the deep learning methods for improving the predictive accuracy in recommendation systems.