SRDPR: Social Relation-Driven Dynamic Network for Personalized Micro-Video Recommendation

With the rapid popularization of the mobile Internet, micro-video has become the mainstream media form today. Compared with traditional video, micro-videos have the characteristics of huge volume, complex video types, and changing user interests. However, micro-video recommendation faces the huge challenge of sparse interaction data, which makes traditional collaborative filtering-based recommendation methods unsuitable for a micro-video recommendation. In addition, only considering the user’s historical behavior for the recommendation while ignoring the user’s social relationship will lead to the problem of a monotonous and cold-start of micro-video recommendation results. To solve the above address, we propose a Social Relation-driven Dynamic network for Personalized micro-video Recommendation, named SRDPR. Specifically, SRDPR divides interest groups by mining users’ public preferences, so active users within the same group can expand user and micro-video interactions. Meanwhile, to perceive users’ dynamic preferences, SRDPR combines a timeline and self-attention mechanism to capture users’ long- and short-term interests and behaviors. Extensive experiments on publicly available datasets show that SRDPR achieves considerable gains over state-of-the-art methods.