Network Effects in Marketing: Graph Neural Network–Based Relationship Prediction and Clustered A/B Testing

Marketing activities that rely on network effects, such as friend‑invited discounts, are difficult to evaluate because user influence can spill over between control and treatment groups, leading to biased conclusions.

To address this, the Ant Intelligent Engine team combined relationship prediction with clustering to improve the precision of activity‑effect estimation. They first predict which users are likely to influence each other using a Graph Neural Network (GNN), then cluster the predicted graph with a distributed Louvain algorithm.

Traditional A/B designs based on cities or raw friend clusters suffer from small sample sizes or weak correlation between friendship and influence. The proposed GNN model aggregates neighbor information at each iteration, incorporates node‑label structural features, and feeds node and edge representations into a fully‑connected MLP to predict invitation likelihood.

After edge prediction, low‑confidence edges are filtered and the weighted graph is clustered using Louvain, which scales to billions of edges (e.g., 10⁹ edges clustered in ~20 minutes, 10¹⁰ edges in ~4 hours). Different clustering strategies—Geo, Louvain, HRLouvain, LinkLabel, LinkLouvain, and LinkLouvain UW—are compared using interference (I) and variance (Var(Y)) metrics derived from KDD‑2018 methods.

Experiments show that the LinkLouvain approach (with edge weight = invitation probability) balances low interference and low variance, completing clustering within six hours on a billion‑node graph. Deployed in Alipay’s large‑scale marketing campaigns, this method achieved conversion‑rate differences within 0.2 % and a 58 % cumulative uplift, enabling rapid, data‑driven decision making.

The team, composed of award‑winning researchers, continues to publish at top data‑mining venues and invites interns to join their efforts.