Graph Algorithms in Alibaba E‑commerce Risk Control: Practices and Insights

This article shares the practice of graph algorithms in Alibaba's e‑commerce risk control system, covering six parts: an introduction to graph algorithms in risk scenarios, graph methods for interaction content risk, product content risk, dynamic heterogeneous graph applications, the ICDM2022 large‑scale risk product detection competition, and a summary with future outlook.

1. Graph Algorithms in E‑commerce Risk Scenarios Alibaba's risk characteristics are highly adversarial and combinatorial, involving diverse markets, business scenes, client platforms, and data sources. Risks are numerous, interrelated, and can shift, making risk control complex.

Graph algorithms enhance adversarial robustness by linking entities (users, accounts, items) to uncover hidden fraud patterns, enabling early detection of coordinated malicious behavior.

2. Interaction Content Risk Using Xianyu comment spam as an example, a heterogeneous graph (items, comments, users) combined with a homogeneous comment graph forms the GAS model. The model processes over 30 k comments, 2 k items, and 9 M users, achieving a 30% increase in recall over previous MLP models.

3. Product Content Risk Two approaches are described: multimodal fusion models that learn product representations from text, images, and metadata, and heterogeneous graph learning that connects products, sellers, and users to improve recall, especially for long‑tail risks.

Graph structure learning for product graphs involves building a K‑NN graph from product embeddings, refining edges via heterogeneous graph transformer (HGT), and iteratively updating embeddings to achieve state‑of‑the‑art performance.

4. Dynamic Heterogeneous Graph Practice Real‑world fraud often follows temporal patterns; dynamic graphs capture these changes. The authors propose an attention‑based dynamic GNN with AutoML (DHGAS) that searches optimal architectures across node‑type, edge‑type, and time dimensions, and a robustness‑focused method (DIDA) that separates essential from non‑essential patterns to mitigate distribution shift.

5. ICDM2022 Competition: Large‑Scale Risk Product Detection The competition provided anonymized large‑scale graph data. Key findings include the benefit of self‑supervised pre‑training, GNN‑based label propagation, and decoupled depth‑width designs.

6. Summary and Outlook The authors summarize deployment best practices (frameworks, semi‑automatic modeling, automated invocation, graph representation as a modality) and outline future work: massive self‑supervised graph representation learning, graph reasoning capabilities, and frequency‑domain and explainability research for dynamic heterogeneous graphs.

References: 1) Spam Review Detection with GCN (CIKM 2019 Best Applied Research Paper); 2) Dynamic Heterogeneous Graph Attention Neural Architecture Search (AAAI 2023); 3) Dynamic Graph Neural Networks Under Spatio‑Temporal Distribution Shift (NeurIPS 2022).

Q&A highlights challenges of low‑homogeneity graphs, extreme class imbalance, and the need for graph federated learning in future security‑critical scenarios.