Source-linked AI summary
CATeye: Coupled Attribute-Topology Invariance Learning for Voucher Abuse Detection
Tian Tian, Shuaicheng Niu, Hao Kuang, Yuanhang Hu, Dong Li, Zhiqi Shen
TL;DR
Voucher abuse detection must handle rapidly evolving patterns and distribution shifts across time and regions, especially when attribute changes induce topology changes. CATeye jointly selects invariant attributes and structure, then uses multi-view objectives to suppress environment-specific variations. Across a proprietary voucher-abuse dataset and the Elliptic benchmark, it consistently outperforms nine baselines and achieves up to an 8.61% average-F1 improvement.
Problem
Voucher abuse patterns shift across time and regions, while coupled attribute-topology shifts allow GNN message passing to amplify non-generalizable signals.
Method
CATeye uses node-adaptive attribute selection, attribute-conditioned invariant-subgraph extraction, and multi-view objectives to emphasize invariant representations.
Results
CATeye consistently outperforms nine strong baselines across the proprietary voucher-abuse and Elliptic datasets, with up to an 8.61% improvement in average F1 score.
Takeaways & Limitations
The results support robust zero-shot generalization to unseen domains without fine-tuning under coupled shifts.
Takeaways & Limitations
The stated deployment setting relies on limited high-confidence labeled training data and targets generalization across time periods and regions.
Abstract
from arXiv · showhide
Voucher abuse poses a major challenge in e-commerce, where malicious users exploit promotional vouchers for profit. Unfortunately, fraud patterns evolve rapidly over time and across regions, causing distribution shifts that degrade existing detection models unless retrained frequently. To tackle this, we propose the Coupled Attribute-Topology Invariance Learning framework (CATeye). The key challenge arises from coupled attribute-topology shift, where edges built from attribute proximity cause environment-driven attribute shift to induce shifted topology, thereby amplifying variant signals through GNN message passing. CATeye sees through such coupled shifts with two learnable selectors. First, an Attribute Invariance Selector (AIS) learns node-adaptive masks to filter out non-invariant attributes. Then, conditioned on retained invariant attributes, an Edge Invariance Selector (EIS) samples an invariant subgraph and isolates non-invariant edges. Using the resulting invariant and non-invariant components, CATeye constructs multiple views and applies view-specific objectives to emphasize domain-invariant representations while suppressing domain-specific variations. Experiments on both a proprietary dataset from Lazada, a major Southeast Asian e-commerce platform, and a public benchmark show that CATeye consistently outperforms nine strong domain generalization and graph anomaly detection baselines, achieving up to an 8.61% improvement in average F1 score over the strongest baseline. Source code is publicly available at https://github.com/Tian0426/CATeye.
1 Introduction
Voucher abuse detection must generalize across rapidly changing time periods and regions, where coupled shifts in attributes and topology amplify non-generalizable signals. CATeye addresses this challenge by jointly isolating invariant attributes and structure and learning from multiple views.
- Voucher abuse creates financial losses and fairness harms through fake accounts, voucher redemption, resale, and seller collusion.
- Abuse evolves over time and differs across regions, causing models trained on one period or country to degrade in new environments.
- Coupled attribute-topology shifts arise because attribute changes affect attribute-based edges, allowing GNN message passing to amplify domain-specific signals.
- Existing graph domain generalization methods often assume independent attribute and structural shifts or enforce invariance only in the embedding space.
- CATeye uses an Attribute Invariance Selector and an Edge Invariance Selector to isolate invariant attributes and an invariant subgraph before multi-view learning.
- CATeye is reported to outperform nine baselines by up to 8.61% in average F1 score across the evaluated datasets.
2 Related Work
Graph anomaly detection methods commonly assume stable train-test distributions, while CATeye targets evolving voucher-abuse patterns by disentangling invariant and variant attributes and graph structure.
- Graph anomaly detection is widely used for fraud, intrusion, and spam detection, including voucher abuse detection through graph-based models.
- Most graph anomaly detection methods implicitly assume that training and test data follow the same distribution.
- Figure 2 depicts CATeye training on k-hop ego-graphs sampled from multiple source graphs and inference on an unseen target graph without fine-tuning.
- CATeye’s selectors disentangle node attributes and graph structure into invariant and non-invariant components that form multiple views.
- The framework explicitly disentangles invariant and variant components in both node attributes and graph structure to address coupled shifts.
3 Methodology
CATeye addresses coupled attribute-topology shifts by selecting invariant attributes and structure, then training on multiple views with objectives that promote invariant predictive representations and suppress variant shortcuts.
- Problem formulation: CATeye formulates voucher abuse detection as graph domain generalization across disjoint source and unseen target graphs without fine-tuning or adaptation.The setup includes distribution shift between source and target graphs and uses localized k-hop ego-graphs as training instances.
- Problem formulation: Coupled shifts arise because environment-dependent attribute and structural variations interact, while attribute-induced edges can propagate non-invariant signals through message passing.The generative description links invariant factors and environment-driven variant factors to observed attributes, structure, and labels.
- Automatic Coupled Invariance Selection: AIS learns node-adaptive binary masks that separate invariant attributes X̂I from non-invariant attributes X̂N.The masks are produced through Gumbel-Softmax reparameterization, and the selected attributes are intended to remain stable across domains.
- Automatic Coupled Invariance Selection: EIS computes edge scores from embeddings based on X̂I, retains selected edges as an invariant subgraph, and separates the remaining non-invariant structure.Using invariant attributes for edge-importance computation prevents non-invariant attributes from disturbing invariant structure selection.
- Multi-View Invariant Representation Learning: The framework combines invariant and non-invariant components into invariant, partially invariant, and two non-invariant views.The invariant view uses X̂I and ÂI; the partially invariant view uses X̂I and ÂN; the two non-invariant views include X̂N.
- Multi-View Invariant Representation Learning: View-specific objectives train the invariant view for label prediction and cross-domain invariance, conditionally constrain the partially invariant view, and maximize entropy on non-invariant views.LCE and LCL are applied to the invariant view, LPI limits reliance on non-invariant structure, and entropy maximization discourages reliance on non-invariant views.
4 Experiments
CATeye is evaluated on a proprietary voucher-abuse dataset and the public Elliptic benchmark under temporal and regional domain shifts. It consistently outperforms strong baselines, while ablations and sensitivity analyses examine its objectives, selectors, and hyperparameters.
- Datasets: The evaluation uses Lazada order graphs spanning Indonesia and Vietnam, campaign and regular days, and a nearly four-month temporal gap, with severe class imbalance.The source domains are ID0501–ID0505, while 16 target graphs cover Indonesia and Vietnam; abusive orders form only a tiny fraction of orders.
- Baselines: The voucher-abuse experiments compare CATeye with ERM, AugAN, four general-purpose domain-generalization methods, and three graph domain-generalization baselines.The comparison includes IRM, IB-IRM, V-REx, DANN, TRACI, EERM, and CaNet.
- Main Results: 81.37% overall average F1 on the voucher-abuse dataset puts CATeye 2.81% above the strongest baseline.CATeye achieves the highest performance across all target domains, with especially evident improvements on campaign days.
- Main Results: 76.84% average F1 on Elliptic exceeds the second-best method by 8.61%, with consistent gains on later test folds T4–T7.These later folds correspond to larger temporal shifts.
- Ablation Study: Disabling any view-specific objective degrades performance, while LPI provides the largest reported enhancement at 5.46%.The ablation study identifies LPI as particularly important for preventing non-invariant structure from contaminating generalization.
- Sensitivity Analysis: CATeye remains stable across a broad range of masking ratios and stays above the second-best baseline without exhaustive ratio tuning.The sensitivity analysis varies the proportions of retained invariant attributes and edges selected by AIS and EIS.
5 Conclusion
The conclusion presents CATeye as a framework for voucher abuse detection under domain shift. Across a real-world dataset and the public Elliptic benchmark, it reports strong zero-shot generalization and outperformance of nine baselines.
- Conclusion: CATeye extracts invariant information from node attributes and graph structure and uses multi-view objectives to reduce reliance on environment-specific variations.The framework is designed for voucher abuse detection under domain shift.
- Conclusion: Experiments on a real-world voucher-abuse dataset and the public Elliptic benchmark show consistent outperformance of nine strong baselines.The conclusion also reports robust zero-shot generalization to unseen domains.
6 GenAI Usage Disclosure
The authors state that large language models were used only to polish writing and improve manuscript clarity. They attribute the research ideas, methods, experiments, analyses, and conclusions to the authors.
- GenAI Usage Disclosure: Large language models were used only for writing polish and manuscript clarity.The disclosure excludes the research ideas, methods, experiments, analyses, and conclusions from LLM involvement.
- GenAI Usage Disclosure: The authors state that they developed and verified all research ideas, methods, experiments, analyses, and conclusions.