Social4Rec: Enhancing Video Recommendation with Social Interest Networks

Introduction The article presents Social4Rec, a video recommendation algorithm that leverages social interest networks to improve recommendation performance, especially for cold‑start users.

Recommendation Landscape Traditional recommendation systems rely on content, knowledge, or collaborative filtering, but they suffer from cold‑start issues for new users and items.

Social Interest Network Social interest is divided into two parts: a social network (direct friendships) and an interest network (user preferences such as favorite videos, stars, or creators). The proposed SocialNet extractor captures these interests and integrates them into a baseline YouTube DNN model via attention.

Social4Rec Architecture The model consists of three components: (1) a coarse‑grained interest extractor (SoNN) using a self‑organizing neural network to assign users to interest groups, (2) a fine‑grained interest extractor based on meta‑path neighborhood aggregation, and (3) integration of the extracted interest vectors into the YouTube DNN model.

Coarse‑Grained Extractor (SoNN) User social features are embedded, then clustered by a self‑organizing neural network. The clusters are refined with K‑Means to merge sparse groups, producing robust interest groups for each user.

Fine‑Grained Extractor (Meta‑Path Aggregation) Within each coarse group, meta‑paths (e.g., User‑Movie‑User) are used to find top‑N similar users; their embeddings are aggregated with the target user’s embedding to form detailed interest vectors.

Interest Vector Fusion The four meta‑path embeddings are combined with the YouTube DNN input via attention, producing a weighted user interest vector that is dot‑producted with item embeddings to compute CTR scores.

Experiments Offline tests show AUC improvement from 0.765 to 0.770 overall and a 2.33 % lift for cold‑start users; online results report a 3.6 % CTR increase overall and 2 % for cold‑start users, along with higher click volume and watch time.

Conclusion Incorporating social interest information significantly boosts recommendation performance, especially for cold‑start scenarios, highlighting the value of multi‑source data integration in modern recommender systems.