Applying Graph Machine Learning in Ant Group's Recommendation System

The presentation introduces the use of graph machine learning in Ant Group's recommendation system, focusing on low‑activity users who have sparse interaction histories. It outlines three sources of auxiliary information—social network UU relationships, knowledge‑graph (II) relations, and other domain UI relations—to enrich user‑item modeling.

Background : Alipay hosts many recommendation scenarios (e.g., waist‑band, fund, coupon) where user behavior is sparse. The goal is to boost DAU by targeting low‑activity users with tailored content.

Graph‑based Recommendation : Knowledge graphs provide rich contextual signals but pose challenges: (1) many irrelevant edges increase training cost; (2) difficulty aggregating user preferences onto graph edges. The solution involves graph distillation and refinement.

Existing Methods : (1) Embedding‑based models pre‑learn graph embeddings but cannot capture user‑specific edge preferences. (2) Path‑based models require extensive expert‑defined meta‑paths and ignore edge‑level user bias. (3) GCN‑based models weight edges by attention independent of target node representations.

Proposed Solution : A four‑stage architecture: (1) Graph representation learning using TransH to map nodes into edge‑specific spaces; (2) Knowledge‑dependency attention to learn edge importance; (3) Distillation module to sample and denoise sub‑graphs; (4) Conditional attention for graph refinement, followed by a dual‑tower model for final prediction. The system jointly optimizes graph‑learning loss and recommendation loss with linear complexity in nodes and edges.

Experiments : Offline and online evaluations on fund‑recommendation and CTR tasks show significant improvements over baselines such as CKE, NMF, RippleNet, and KGAT. Ablation studies confirm the effectiveness of conditional and knowledge attention components.

Social and Text‑based Recommendation : To aid operations, a hybrid GNN‑Topic‑Model approach incorporates UU and UI relations, using a Logistic‑Normal prior for user‑intent distributions and a skip‑gram‑based text encoder, achieving gains in both offline similarity metrics and online A/B tests.

Cross‑Domain Recommendation : A CD‑GNN aligns feature spaces of active and inactive users via domain‑invariant layers, enabling transfer of click signals from active to dormant users. Experiments demonstrate superior CTR performance compared to standard GCNs, even under severe sparsity.

Q&A : The authors clarify that CD‑GNN layers are not shared due to differing feature distributions, suggest pre‑training or feature reconstruction for scarce target labels, and note that the graph‑based model is deployed in a re‑ranking stage with typically two‑hop GNNs for latency reasons.

The talk concludes with acknowledgments and a call for audience interaction.