COUSIN: A network-based regression model for personalized recommendations

摘要

Recently, such state-of-the-art methods as collaborative filtering, content-based, model-based and graph-based approaches have achieved remarkable success in recommendations. However, most of them make recommendations based on either information from users or objects, or bipartite relationships between them, without explicitly exploring object, user and object-user relationships simultaneously. Meanwhile, recent discoveries in sociology and behavior science have demonstrated that similar users tend to select similar objects, usually referred to the n-degree of influence. However, such understandings have not been systematically incorporated into recommendations yet. With these understandings, we propose a novel method named COUSIN (Correlating Object and User SImilarity profiles to personalized recommendatioN), adopting a regression model to incorporate object, user and object-user associations simultaneously in a global way for personalized recommendation. We also construct a power-law adjusted heterogeneous network for COUSIN to prevent adversely influence of popular nodes. We demonstrate the effectiveness of our method through comprehensive cross-validation experiments across two data sets (MovieLens and Netflix). Results show that our method outperforms the state-of-the-art methods in both accuracy and diversity performance, indicating its promising future for recommendation.