Leveraging heterogeneous information networks for personalized entity recommendation

Recommender systems;Entity recommendation;Heterogeneous information networks

Norick, Brandon ; Han ; Jiawei

关键词: Recommender systems;    Entity recommendation;    Heterogeneous information networks;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/99227/NORICK-THESIS-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Recommendation is a challenging but important task which has applications in nearly every sector of industry as well as in academia. There are a wide variety of approaches to the recommendation problem, with network-based techniques garnering increasing interest and study in recent years. However, most of these studies only explore the problem in the context of a single relationship between entities, such as a following relationship in a social network like Twitter. Such approaches ignore the complex environment in which most recommendation tasks exist in favor of simplifying the problem. The complexity of human decision making necessitates approaches which can utilize the heterogeneous environments in which the recommendation task is set rather than reducing them to single relationship.In this work, we explore the problem of entity recommendation without such a simplification, instead we utilize heterogeneous information networks to capture the complexity of the behaviors for which we are seeking to make recommendations. Our proposed approach captures the different behaviors of individuals by examining their heterogeneous relationships in the network and as a result can provide high-quality personalized recommendations from implicit feedback represented in heterogeneous information networks.We begin by introducing meta-path-based latent features, which capture the connectivity of entities in the network along different paths, giving us a foundation which explicitly accounts for the heterogeneous nature of the task. Upon this foundation we build a global recommendation model using a ranking optimization technique known as Bayesian Personalized Ranking. We extend this global model into a personalized model, building a model which can capture the differences present in the network that describe the preferences of different users. Finally, empirical studies show that our techniques are more effective than several popular and state-of-the-art entity recommendations techniques.