Social4Rec: Distilling User Preference from Social Graph for Video Recommendation in Tencent (2302.09971v3)

Abstract: Despite recommender systems play a key role in network content platforms, mining the user's interests is still a significant challenge. Existing works predict the user interest by utilizing user behaviors, i.e., clicks, views, etc., but current solutions are ineffective when users perform unsettled activities. The latter ones involve new users, which have few activities of any kind, and sparse users who have low-frequency behaviors. We uniformly describe both these user-types as "cold users", which are very common but often neglected in network content platforms. To address this issue, we enhance the representation of the user interest by combining his social interest, e.g., friendship, following bloggers, interest groups, etc., with the activity behaviors. Thus, in this work, we present a novel algorithm entitled SocialNet, which adopts a two-stage method to progressively extract the coarse-grained and fine-grained social interest. Our technique then concatenates SocialNet's output with the original user representation to get the final user representation that combines behavior interests and social interests. Offline experiments on Tencent video's recommender system demonstrate the superiority over the baseline behavior-based model. The online experiment also shows a significant performance improvement in clicks and view time in the real-world recommendation system. The source code is available at https://github.com/Social4Rec/SocialNet.

Incorporating Social-aware User Preference for Video Recommendation

This paper presents a novel approach to video recommendation systems by introducing Social4Rec, a framework that leverages social-aware user preferences to augment existing behavior-based models. The authors address significant challenges in current recommendation systems, such as user behavior sparsity and the cold user problem, by utilizing a social graph to refine interest modeling.

The core contribution of this work is the integration of a social graph, which encapsulates multiple relation types among users to enhance interest pattern recognition in recommendation models. Traditional behavior-based models often suffer due to sparse interaction data and inherent noise from user experiences that don't accurately capture interest complexity. Social4Rec proposes a robust solution to this limitation by implementing a Cluster-Calibrate-Merge (CCM) network. This network is designed to identify users with similar interests through an enriched understanding of social connections and nuanced relational data.

The CCM network operates through three layers: a cluster layer based on self-organizing neural networks, a calibrator layer that uses knowledge distillation for robustness, and a merge layer to finalize interest group assignments. By capturing latent interest patterns and mitigating the effect of sparse or trivial social relations, Social4Rec is able to enhance the user representation derived from behavior data with a refined social embedding.

In terms of computational performance, Social4Rec demonstrates superiority in both offline and online contexts. Offline experiments show a substantial improvement in AUC, especially in cases involving cold users with limited historical data. These users typically provide the most challenging test cases for recommendation systems. Online experiments on a global video recommendation platform further substantiate the efficacy of the model, revealing significant increases in CTR and user engagement metrics such as click numbers and view time. The results highlight the potential of Social4Rec to improve user modeling by utilizing social data—ensuring recommendations are more personal and relevant.

From a theoretical perspective, this paper introduces a paradigm shift towards using multi-source data aggregation for refining user profiles. By emphasizing a holistic view of user data through social interactions, the research enhances the theoretical framework for tackling fundamental challenges in recommendation systems. Practically, it opens pathways for the development of more intelligent systems that can predict and adapt to user needs more accurately.

Looking forward, the broad application of Social4Rec suggests a permutation in how recommendation systems can be designed—moving towards increasingly personalized, data-enriched environments. Future research may explore extending this framework to different domains beyond video recommendation and incorporating even richer interaction data from user activity streams. Additionally, expanding the model to dynamically adjust to evolving social connections could further enhance its applicability, ensuring that recommendation systems keep pace with rapidly changing user behaviors and interests.