Community Recommendation in Tencent Games: Adaptive K‑Free Community Detection and Constrained Large‑Scale Community Recommendation (ComRec)

The talk introduces community recommendation in Tencent games, starting with a definition of social networks, the concept of communities, and why recommending communities can boost player activity and revenue by alleviating information overload and cold‑start issues.

Motivation is illustrated with figures showing higher playtime and payment rates for players belonging to communities versus those without, emphasizing the need for data‑driven methods to select suitable communities for each player.

Two core research problems are addressed: (1) adaptive K‑Free community detection to discover latent groups without knowing the number of communities, and (2) a constrained large‑scale community recommendation algorithm (ComRec) that respects the rule that a player can belong to only one community at a time.

Adaptive K‑Free Community Detection (DAG) combines a Mask AutoEncoder (MAE) for node embedding with a Community Affiliation Network (CAN) that makes community assignment differentiable, allowing joint optimization of modularity (structural cohesion) and semantic similarity. Sparse constraints on the read‑out matrix enable automatic determination of the appropriate number of communities.

The proposed EDGE metric evaluates community quality by balancing structural and semantic aspects, overcoming limitations of traditional modularity or pure semantic scores.

Constrained Large‑Scale Community Recommendation (ComRec) introduces a labeling mechanism to indicate whether a node belongs to any community, a graph neural network for user‑item representation, and a two‑stage propagation scheme (global column‑wise and local subgraph‑wise) that scales to graphs with tens of millions of nodes and edges while preserving efficiency.

Extensive offline experiments on four Tencent game social graphs (millions of nodes, tens of millions of edges) compare DAG and ComRec against baselines such as Louvain, Deep Graph Clustering, LightGCN, GraphRec, and MLP‑based models, showing 3.5%–5% improvements in hit‑rate and NDCG. Online A/B tests on a 20 M‑node graph with 280 k communities demonstrate 4%–9% lift in exposure‑through‑rate and click‑through‑rate, with the proposed models also achieving the lowest inference latency.

The Q&A section clarifies that the first algorithm is unsupervised, that the methods can be adapted to heterogeneous graphs by swapping the encoder, and discusses the bias between pre‑training and recommendation tasks.

Finally, the team behind the work is introduced: the Tencent Interactive Entertainment CDP Data Science Center Social Algorithm Group, focusing on graph‑based social recommendation, friend/party matching, influence maximization, and social marketing for games.