Enhancing Collaborative Filtering-Based Rating-Prediction by Discovering and Incorporating User Concerns from User Review

Collaborative filtering (CF) based rating prediction discover similar users and items to predict unknown ratings based on how similar users have rated similar items. Discovering and infusing additional knowledge to supplement the discovery of similar users can potentially improve the accuracy of CF-based rating prediction. Neighborhood discovery and non-negative matrix factorization (MF) are very popular techniques for collaborative filtering. Hence, we try to improve these two techniques to increase rating prediction accuracy by using multi-view clustering to better discover neighborhoods and infusing additional user behavior knowledge, respectively. Additional user behavior knowledge is extracted by mining User-Concern vectors (UC-vectors) which represent hierarchical relationships between hidden user concerns in user reviews. To further improve the rating prediction accuracy, we incorporate the extracted UC- vectors into Deep Neural Network (DNN) based architectures. We propose and experiment with a DNN architecture consisting of two parallel branches to learn user and item latent vectors before aggregating them to make the final rating predictions. The initial input to our DNN are the user and item rating behavior vectors constructed by collecting all the ratings given or received by the users or the items, respectively. We then regulate the learning of the latent user vectors using UC-vectors. In addition to using only rating behavior vectors as the initial input to our DNN model, we also explored using Term Frequency--Inverse Document Frequency (TF-IDF) vectors as the input to DNN. However, TF-IDF is based on statistical learning alone and does not directly capture the conceptual contents of the text or the behavioral aspects of the writer. Hence, we also propose a novel weighing scheme to extract relatively low dimensional user behavioral vectors similar to our previously extracted UC-vectors and append them to the corresponding user or item TF-IDF vectors. Such additional user behavioral knowledge will allow TF-IDF to better capture user similarity and to further improve rating predictions. Our experiments and results on standard neighborhood-based, NMF-based and DNN-based architectures clearly showed that infusing additional user behavioral aspect can significantly improve the rating prediction accuracy.