Research map / Graph learning / Knowledge and temporal graphs
Dynamic Graph Learning with Static Relations for Credit Risk Assessment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DGNN-SR, which combines a dynamic user-merchant transaction graph (with a multi-view time encoder and temporal attention) and a static user-user fund transfer graph via adaptive re-weighting, for credit risk assessment.
Problem
Credit default detection on online payment platforms; existing graph methods ignore user-to-user fund transfers and underuse temporal information (only relative time).
Approach
Models user-merchant payments as a continuous-time dynamic graph and user-user transfers as a static graph, encodes both relative and absolute/semantic time with a multi-view time encoder, adds time-aware temporal attention, and fuses the two branches with a self-gated re-weighting unit.
Key results
- DGNN-SR variants reach AUC 0.8215 (DyGFormer-SR) on D1 and 0.8553 (DyGFormer-SR) on D2, versus 0.8046 for best dynamic baseline GraphMixer on D1 and 0.8468 for GraphSAGE on D2.
- Abstract claims 0.85%-2.5% improvement over existing SOTA methods.
- Re-weighting fusion outperforms simple concatenation in most backbone combinations (e.g., GraphSAGE + DyGFormer +0.22% AUC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine dynamic and static graph representations with DGNN-SR | Dynamic GNN with static user-user graph · D1 | AUC: 0.8200 ± 0.0006 -> 0.8209 ± 0.0008 | Helped |
| Combine dynamic and static graph representations with DGNN-SR | Dynamic GNN with static user-user graph · D2 | AUC: 0.8546 ± 0.0013 -> 0.8553 ± 0.0016 | Helped |
| Add a trainable multi-view time encoder (ablation) | TGAT and DyGFormer · D1 and D2 | AUC | Helped |
| Use the fixed time encoder for relative and absolute time (ablation) | GraphMixer · D1 and D2 | AUC | Mixed |
| Add temporal attention to the MTE-enhanced dynamic GNN (ablation) | TGAT and DyGFormer · D1 | AUC | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | TGAT + GCN · D1 | AUC: 0.8070 ± 0.0012 -> 0.8082 ± 0.0012 | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | TGAT + GAT · D1 | AUC: 0.8049 ± 0.0011 -> 0.8052 ± 0.0019 | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | TGAT + GraphSAGE · D1 | AUC: 0.8178 ± 0.0009 -> 0.8181 ± 0.0006 | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | DyGFormer + GCN · D1 | AUC: 0.8136 ± 0.0012 -> 0.8134 ± 0.0005 | Hurt |
| Use re-weighting instead of concatenation for fusion (ablation) | DyGFormer + GAT · D1 | AUC: 0.8176 ± 0.0008 -> 0.8197 ± 0.0009 | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | DyGFormer + GraphSAGE · D1 | AUC: 0.8185 ± 0.0009 -> 0.8207 ± 0.0005 | Helped |
| Use re-weighting instead of concatenation for fusion (ablation) | GraphMixer + GCN · D1 | AUC: 0.8136 ± 0.0007 -> 0.8155 ± 0.0008 | Helped |
Surprising findings
- Fixed (non-trainable) time encoder in GraphMixer gave no gain on D1 and degraded AUC on D2, while trainable multi-view encoding helped TGAT and DyGFormer.
Limitations
- Evaluated only on two proprietary Tencent Mobile Payment datasets, which cannot be released.
- Gains over strongest baselines are small (roughly 0.1-0.3% AUC in many settings).
Experimental setup
Datasets: D1 (Tencent Mobile Payment) D2 (Tencent Mobile Payment)
Benchmarks: D1 D2
Metrics: AUC Recall@10% TP@10%
Baselines: GCN GAT GraphSAGE TGAT DyGFormer GraphMixer concatenation fusion of dynamic and static backbones
Builds on: TGAT GraphMixer DyGFormer GraphSAGE
Closest papers by meaning
- MDGNN: Multi-Relational Dynamic Graph Neural Network for Comprehensive and Dynamic Stock Investment Prediction (AAAI 2024 · similarity 0.85)
- Role Perceptual Augmented Temporal Graph Network for Related-party Transaction Detection (AAAI 2026 · similarity 0.84)
- Temporal-Aware Evaluation and Learning for Temporal Graph Neural Networks (AAAI 2025 · similarity 0.84)
- Semi-supervised Credit Card Fraud Detection via Attribute-Driven Graph Representation (AAAI 2023 · similarity 0.83)
- Multi-Scale Dynamic Graph Learning for Time Series Anomaly Detection (Student Abstract) (AAAI 2024 · similarity 0.82)
- DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.