A personalized paper recommendation method considering diverse user preferences

Prior studies of paper recommendation methods that consider historical user preferences rarely adequately address the complexity of user preferences and interests. We propose a method to recommend personalized papers based on a heterogeneous network that includes papers, venues, authors, terms, and users as well as the relations among these entities. We investigate meta-paths in the network to capture user preferences and apply random walks on these meta-paths to measure recommendation scores of candidate papers to target users. We employ a personalized weight learning process to discover a user's personalized weights on different meta-paths using Bayesian Personalized Ranking as the objective function. A global recommendation score is calculated by combining recommendation scores on different meta-paths with personalized weights. We conducted experiments using two different datasets and the results showed that the proposed method performed better than other baseline methods.