Empowering Credit Risk Detection in Weixin Pay with Billion-Scale Deep Graph Learning

Summary

This paper introduces RAOS (Risk-Aware Overlapping Subgraph learning), a scalable graph learning framework designed for billion-scale credit risk detection in Weixin Pay. The authors state: To bridge the gap between industrial scalability and high-precision risk detection, we propose a Risk-Aware Overlapping Subgraph Learning framework (RAOS).

The core problem addressed is that deploying GNNs in an industrial setting faces a severe scalability bottleneck. Real-world graphs typically scale to billions of nodes and edges. The authors note that existing subgraph-based distributed training strategies often compromise topological integrity for load balancing. This can be catastrophic for risk detection, as it indiscriminately severs the long-tail evidence chains essential for risk propagation.

The proposed framework consists of three main components:

1. Risk-Aware Subgraph Construction: The authors first decompose the full graph into balanced base partitions to ensure computational load balance using random hash partitioning. Then, they perform a budget-constrained sampling approach that selectively replicates informative long-tail nodes while filtering out noise, thereby preserving critical risk diffusion patterns and avoiding redundancy. The selection is based on the h-index: We prioritize the tail nodes to augment the local subgraph. Specifically, we first filter candidates with lower h-index scores (h-index ≤ θ) and sample b nodes to obtain the final expansion set.

2. Dynamic Heterogeneous Graph Learning: For each subgraph, the authors propose a Dynamic Heterogeneous Graph Encoder (DHGNN) as the backbone. This encoder utilizes a Heterogeneous Graph Transformer (HGT) to capture spatial patterns, then incorporates an attention mechanism and GRU to model temporal dynamics. The final spatio-temporal node representation is computed as: H t = GRU(h t, H̃ t−1).

3. Cross-Subgraph Consistency Alignment: To address representation inconsistency, the authors introduce a cross-subgraph consistency alignment module that harmonizes local views into a globally consistent latent space using node-level Canonical Correlation Analysis (CCA). The loss consists of an invariance term and a decorrelation term: L CCA = L inv + λ n L dec.

The model is trained with a joint objective: L total = L BCE + λ CCA L CCA, where L BCE is the binary cross-entropy loss on labeled users.

Experiments were conducted on WeCreditFraud, a large-scale real-world dataset provided by Tencent, sourced from the Weixin Pay platform containing over 845M user behavior records involving approximately 498M unique users and covering over 60 distinct types of risk events. The authors evaluated both current risk (labels at observation date) and forward risk (labels over a 90-day forward window).

Key results show RAOS consistently achieves superior performance compared to all baselines for credit risk detection. Specifically, Compared to Random Hash, which severs risk paths, RAOS achieves a KS improvement of over 10.5%. The framework also outperforms the full-graph training backbone, which the authors attribute to the structural denoising effect of the long-tail overlapping strategy.

Ablation studies confirm that the overlapping strategy is essential, valid risk patterns are concealed in long-tail entities, and representation consistency is critical. Parameter sensitivity analysis shows that an appropriate selection of the overlap ratio represents an optimal trade-off and that limiting selection to low h-index nodes forces the model to focus on informative long-tail nodes.

For system deployment, the authors implemented a training system based on the distributed data parallel paradigm and report that RAOS trains on WeCreditFraud on 8 H20 GPUs in 16 hours, with peak memory 66.6 GB per worker, communication and I/O overhead at 5.5% and 0.75% per epoch, and inference latency of 2.78 s for 1M users.

Online A/B tests ran for 30 days with a 50/50 traffic split between the baselines and RAOS. Results show "the most significant breakthrough is observed in the inactive segment. Compared with Baseline-v2, RAOS achieves a remarkable performance leap, increasing the user-level KS by 4.06% and the amount-weighted KS by 5.41% on inactive user risk detection. The authors also report that RAOS improves both metrics across all risk segments, with all gains being statistically significant (p < 0.01)."

The authors conclude: "Extensive offline evaluations and online deployment on billion-scale graphs demonstrate the effectiveness of RAOS as a scalable instrument for credit risk detection. It helps minimize financial losses and prevent the propagation of credit fraud risks, contributing to a stable financial ecosystem."

Based on the paper, here are the specific improvements I can make to AI systems, along with the resulting capabilities:

Improvement: Replace standard disjoint graph partitioning (e.g., random hash, METIS) with a budget-constrained, h-index-filtered overlapping sampling strategy. This selectively replicates long-tail nodes (h-index ≤ threshold) across subgraphs while filtering out high-degree hubs that introduce noise.

Resulting Capability: The AI system can preserve critical risk diffusion paths that would otherwise be severed by partitioning. It achieves a 10.5% KS improvement over random hash partitioning on current risk detection and a 5.86% KS improvement on forward risk prediction, while maintaining computational load balance.

Improvement: Add a Canonical Correlation Analysis (CCA)-based alignment loss that enforces both invariance (same user, same representation across subgraphs) and decorrelation (redundancy reduction across feature dimensions) on overlapping nodes.

Improvement: Integrate a Heterogeneous Graph Transformer (HGT) with edge-type-aware attention and a GRU-based temporal module that uses attention-weighted historical states (window size = 3).

Improvement: Replace random or high-degree node expansion with low h-index node selection, effectively filtering out structural noise from the full graph while amplifying sparse risk signals.

1. Detect credit fraud at billion-scale (498M users, 845M records) with 52.96 KS / 81.57 AUC on current risk and 26.47 KS / 65.72 AUC on 90-day forward risk, outperforming all baselines including full-graph training.

2. Identify latent risk users who appear normal at observation time but become fraudulent within 90 days, with a 5.86% KS improvement over disjoint partitioning methods.

3. Operate efficiently in production with:

• Training on 8 GPUs in 16 hours for a billion-scale graph

• Peak memory of 66.6 GB per worker

• Communication overhead of only 5.5% per epoch

• Inference latency of 2.78 seconds for 1M users

4. Improve financial loss prevention in online deployment:

• 4.06% user-level KS improvement and 5.41% amount-weighted KS improvement on inactive users (previously hard-to-detect segment)

• Up to 7.5% improvement in risk amount recall across all risk segments

• 4.9% improvement in risk user recall in the top 5-20% risk score interval

5. Generalize across different GNN backbones (GAT, GraphSAGE), consistently improving KS by 0.70–1.01% on forward risk prediction, making it a drop-in scalable solution for existing graph learning systems.

6. Maintain robust performance across hyperparameter variations (overlap ratio 30–50%, h-index threshold 1–2), providing stable deployment without extensive tuning.

Credit risk detection, particularly mitigating individual fraud, is crucial for maintaining the stability of digital financial ecosystems. Accurately identifying credit fraud among billions of users is critical for minimizing financial losses and safeguarding the sustainability of inclusive financial services. Given that credit fraud risks are often concealed within heterogeneous user-risk graphs, Graph Neural Networks (GNNs) have emerged as an effective tool for risk mining by capturing complex dependencies. To address the scalability bottleneck of industrial GNNs, distributed training based on subgraphs is indispensable. However, existing strategies often compromise topological integrity for load balancing. This can be catastrophic for risk detection, as it indiscriminately severs the long-tail evidence chains essential for risk propagation. Overlapping subgraphs can restore severed risk contexts but inevitably introduce redundancy and noise, while overlooking the representation alignment across different local subgraphs. In this paper, we propose a risk-aware overlapping subgraph learning framework for large-scale credit risk detection. We first construct base partitions to ensure load balance. Then, we perform budget-constrained sampling that selects informative long-tail nodes, thereby preserving critical risk diffusion patterns while filtering out noise. To mitigate representation inconsistency, we design a cross-subgraph consistency alignment mechanism. By enforcing alignment constraints on the overlapping nodes, we harmonize the local representations into a globally consistent latent space. Extensive experiments on Weixin Pay's production dataset demonstrate that our model significantly outperforms existing strategies for risk detection, offering a scalable and effective solution for industrial graph learning.