Infusing Latent User-Concerns from User Reviews into Collaborative Filtering

Traditionally, Collaborative Filtering (CF) based recommendation employs past rating behaviors of users on items to discover similar users and similar items. We can further improve on discovering user similarities by better understanding user behaviors through analyzing user reviews. In their reviews, users generally mention about things that are of greater interest to them, and these cues can provide an effective medium to discover users with similar interests and concerns. In this paper, we extract latent User-Concerns from user reviews and construct their hierarchical tree (UC-Tree). By associating each user with the corresponding concerns in the UC-Tree, we then generate vectors that represent intricate user behaviors. Finally, we infuse such additional knowledge about the users into the conventional CF-based rating prediction process. Our experiments and results show that such additional behavioral knowledge assists the discovery of similar users and improves the accuracy of conventional CF-based rating prediction.