Predicting User Preferences via Heterogeneous Information Network and Metric Learning

Abstract

Heterogeneous information network contains richer semantic information, considering multiple types of objects and relationships to more accurately determine user preferences. In addition, existing approaches usually project each user to a point in the space, it is insufficient to accurately model the intensity of the user-item relationship and the heterogeneity of different types of objects and their relationships in implicit feedback. In order to solve these problems, we propose Predicting User Preferences via Heterogeneous Information Network and Metric Learning (PUHML). First, we use heterogeneous information networks to model complex heterogeneous data, and obtain users and item node representations. Second, we construct a user-item relationship vector, and translate each user toward items according to the user-item relationship. Finally, to alleviate the limitation of inner product as a scoring function, we introduce metric learning instead of dot product, and use distance to measure user preferences. Experimental results on three datasets demonstrates the effectiveness of our proposed approach over some competitive baselines.