Behavioral graph fraud detection in E-commerce

Hang Yin, Zitao Zhang, Zhurong Wang, Yilmazcan Özyurt, Weiming Liang, Wenyu Dong, Yang Zhao, Yinan Shan

Abstract

Traditional fraud detection models often struggle with sparse or noisy "hard links" like IP addresses or device IDs. This paper presents a novel framework that utilizes behavioral biometric similarities—"soft links"—to construct transaction graphs. By embedding user navigation patterns and applying unsupervised Graph Neural Networks (GNNs), the system achieves significant precision gains, particularly in cold-start scenarios for new and guest buyers.

Introduction

In the evolving landscape of e-commerce, fraud detection has become a sophisticated arms race. Modern fraudsters often operate in organized "fraud rings," making individual transactions difficult to flag using static features alone. While Graph Neural Networks (GNNs) offer a way to detect these collective signals, their effectiveness is heavily dependent on the quality of the underlying graph construction.

Most existing graph-based models rely on "hard links"—direct connections between transactions sharing specific entities like credit card numbers, device IDs, or IP addresses. However, these identifiers suffer from two primary issues:

  1. Sparsity: Strong identifiers (e.g., device IDs) are often unique to a single user, resulting in a sparse graph that prevents effective information propagation.
  2. Noise and Hubs: Weak identifiers (e.g., public IP addresses) can create massive "hubs," connecting thousands of unrelated users and making it nearly impossible to distinguish legitimate users from malicious actors.

To overcome these limitations, researchers from eBay China and ETH Zürich proposed a shift toward behavioral "soft links." This approach treats a user's navigation pattern as a unique biometric signature.

The high-level four-stage workflow of the behavioral graph fraud detection system.

1. Behavioral Embedding

The first step involves converting raw clickstream data into dense vector representations. Using an unsupervised seq2item approach, the team trained an attention-based biLSTM model to process user page-view sequences (e.g., Sign In → View Item → Add to Cart). This captures the "rhythm" and intent of a user's session.

The process of converting behavior and time sequences into dense clickstream embeddings using biLSTM.

2. Efficient Graph Construction

To manage the computational cost of comparing millions of transactions, the framework employs an in-house GPU-based clustering algorithm called pHDBSCAN. By clustering embeddings first, the system can efficiently identify similar behaviors and prune "hub" nodes that would otherwise introduce noise into the graph.

3. GNN Feature Extraction

Once the soft-link graph is built, GraphSAGE (an inductive GNN) is applied to aggregate features from behavioral neighbors. These aggregated "boosting features" are then combined with traditional statistical features in a final Gradient Boosted Decision Tree (GBDT) model for fraud classification.

Key Results and Insights

The behavioral graph approach demonstrated significant improvements, particularly in the "cold start" problem involving new or guest buyers where historical data is absent.

  • Precision Lift: For guest buyers, the model increased precision from 0.82 to 0.86 at a constant recall rate.
  • Efficiency Gains: The use of pHDBSCAN reduced graph-building runtime from several days to just five minutes, making the system viable for large-scale industrial applications.
  • Robustness: Navigation patterns are significantly harder for fraudsters to spoof than hardware identifiers or IP addresses, providing a more resilient foundation for security.

Conclusion

By moving from rigid physical identifiers to fluid behavioral signatures, this research provides a robust framework for detecting sophisticated fraud rings. The transition to "soft links" allows for a more nuanced and effective representation of user intent, paving the way for the next generation of e-commerce fraud detection.

Summary updated

BibTeX
@inproceedings{yin2022behavioral,
  author = {Hang Yin and Zitao Zhang and Zhurong Wang and Yilmazcan Özyurt and Weiming Liang and Wenyu Dong and Yang Zhao and Yinan Shan},
  title = {{Behavioral graph fraud detection in E-commerce}},
  booktitle = {Workshops @ ICDM},
  year = {2022},
  doi = {10.1109/icdmw58026.2022.00105},
}
Figure