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

The presentation introduces the topic of community recommendation in Tencent games, outlining four parts: an overview of community recommendation, an adaptive K‑Free community detection algorithm (published at KDD 2024), a constrained large‑scale community recommendation algorithm (ComRec, published at KDD 2023), and a brief team introduction.

Community recommendation aims to increase player activity and payment by suggesting suitable in‑game communities, addressing information overload caused by millions of existing groups and improving user engagement through social interactions.

Challenges include extreme sparsity of user‑community interactions (cold‑start) and the massive scale of game social graphs (hundreds of millions of nodes and edges). Traditional methods either ignore semantic attributes or require a known number of communities.

The adaptive K‑Free community detection algorithm (DAG) solves these issues by jointly learning structural and semantic information without a prior community count. It uses a Mask AutoEncoder for node embeddings, a Community Affiliation Network for differentiable community selection, and sparsity constraints (L1 on rows, L2 on columns) to produce a high‑dimensional sparse community ID vector.

DAG’s training includes unsupervised feature reconstruction and link‑prediction (BPR) losses, enabling it to discover latent communities and alleviate cold‑start problems.

The constrained large‑scale recommendation algorithm ComRec introduces a labeling mechanism where each node can belong to at most one community, and employs global and local feature propagation to efficiently handle graphs with tens of millions of nodes. Global propagation reduces complexity to sub‑linear time, while local propagation refines community‑level embeddings.

Extensive offline experiments on four Tencent game social graphs (millions of nodes) show ComRec improves hit rate and NDCG by 3.5%–5% over baselines such as LightGCN, GraphRec, and traditional network embedding methods. Online A/B tests on a graph with 20 million nodes, 280 k communities, and 1.6 billion edges demonstrate 4%–9% lift in exposure conversion and click‑through rates, with lower latency than competing models.

The team also proposes a new evaluation metric, EDGE, which balances modularity and semantic similarity for community quality assessment.

Finally, the presenters introduce their team, the Tencent Interactive Entertainment CDP Data Science Center Social Algorithm Group, and invite interested researchers to explore their open‑source code and collaborate on future social algorithm research.