Social4Rec: Social Interest Enhanced Video Recommendation Algorithm

The article presents Social4Rec, a video recommendation algorithm that leverages social interest information to address the user cold‑start problem.

It begins with an overview of the current state of recommendation systems, noting the prevalence of information overload and the evolution from simple content‑based and collaborative‑filtering methods to deep‑learning‑driven approaches, while highlighting the persistent cold‑start challenge.

Social interest networks are introduced, comprising two components: a social network capturing direct user‑to‑user relationships (e.g., friends on Facebook, Twitter) and an interest network reflecting users' affinities to content categories such as movies, celebrities, or favorite creators.

Social4Rec consists of three main parts:

Coarse‑grained interest extractor (SoNN) that uses a self‑organizing neural network to cluster user embeddings into interest groups, followed by K‑means aggregation to merge sparse groups.

Fine‑grained interest extractor based on meta‑path neighborhood aggregation, which selects top‑N similar users via meta‑paths (e.g., User‑Movie‑User) and aggregates their embeddings.

Integration of the extracted interest vectors with a baseline YouTube DNN model through attention mechanisms, producing a final user representation used for CTR prediction.

The coarse extractor maps user side information (age, gender, etc.) into embeddings, assigns them to interest groups, updates group weights using distance‑based formulas, and refines the groups through iterative training and K‑means clustering.

The fine extractor defines meta‑paths such as UMU (User‑Movie‑User) to find related users, aggregates the top‑N neighbor embeddings, and combines four types of interest vectors (movie, celebrity, creator, friend) with the YouTube DNN embedding via attention.

Interest vector aggregation concatenates the user’s base features with the attention‑weighted interest embeddings, passes them through an MLP, and computes the final CTR score by inner‑product with item embeddings.

Experimental results on a 15‑day online traffic log show that Social4Rec improves offline AUC from 0.765 to 0.770 for all users and yields a 2.33‑point gain for cold‑start users; online CTR increases by 3.6% overall and 2% for cold‑start users, with notable boosts in click volume and watch time.

The summary emphasizes that incorporating social interest signals provides a practical and effective way to enhance recommendation performance, especially for cold‑start users.

Q&A sections address meta‑path selection, determining the number of clusters, and why the model benefits cold‑start users.

Additional information mentions the OneRec series of algorithms, open‑source code, and related research papers.