KPG4Rec: Knowledge Property-Aware Graph for Recommender Systems

The collaborative filtering (CF) based models have the powerful ability to use the interaction of users and items for recommendation. However, many existing CF-based approaches can only grasp the single relationship between users or items, such as item-based CF, which utilizes the single relationship of similarity identified from user-item matrix to compute recommendations. To overcome these shortcomings, we propose a novel approach named KPG4Rec which integrates multiple property relationships of items for personalized recommendation. In the initial step, we extract properties and corresponding triples of items from an existing knowledge graph, and utilize them to construct property-aware graphs based on user-item interaction graphs. Then, continuous low-dimensional vectors are learned through node2vec technology in these graphs. In the prediction phase, the recommendation score of one candidate item is computed by comparing it with each item in the user history preference sequence, where the pretrained embedding vectors of items are used to take all the properties into consideration. On the other hand, Locality Sensitive Hashing (LSH) mechanism is adopted to generate brand new preference sequences of users to improve the efficiency of KPG4Rec. Through extensive experiments on two real-world datasets, our approach is proved to outperform several widely adopted methods in the Top-N recommendation scenario.