Modeling User Propagation Ability for Social Recommendation and Influence Maximization in Games

The presentation shares research from NUS and Tencent on social diffusion and influence‑maximization algorithms applied to gaming scenarios. It first describes how user propagation ability metrics are used to improve familiar‑friend recommendation by modeling invitation behavior and incorporating a conversion‑funnel concept into the diffusion process.

A conversion‑funnel model is introduced, extending the classic Independent Cascade (IC) model with probabilities P uv , β, and γ to capture invitation, acceptance, and re‑invitation stages. The model is simulated via Monte‑Carlo methods, estimating spread and enabling capacity‑constrained influence maximization.

Four experiments (two offline, two online) evaluate cascade estimation and diffusion prediction on six datasets (four internal Tencent invitation logs and two public datasets). The proposed ICI method consistently achieves the lowest RMSE and superior AUC/MAP compared to baselines such as Degree and PageRank.

For game‑specific viral marketing, the work adapts influence maximization by selecting seed users under capacity limits, using greedy algorithms (standard greedy and round‑robin greedy) and a scalable implementation (RR‑OPIM+). The approach attains near‑optimal (1/2‑ε) approximation with near‑linear runtime.

Extensive offline and online evaluations show that RR‑OPIM+ and its variants outperform traditional methods in spread, running time, and real‑world activity promotion metrics, confirming the practical value of the proposed diffusion models and algorithms in gaming platforms.

The talk concludes with a Q&A covering probability modeling, evaluation metrics, and mitigation of network effects in AB testing, emphasizing the relevance of these techniques for large‑scale social recommendation systems.