Research map / Continual learning / Online and replay-based
Pre-trained Online Contrastive Learning for Insurance Fraud Detection
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 POCL, a GNN for online medical insurance fraud detection that pre-trains a GraphSAGE encoder with a contrastive-style loss, then fine-tunes it with a GAT detector and updates it online with a Temporal MAS importance-weight penalty.
Problem
Medical insurance fraud patterns evolve over time, so static offline fraud detectors degrade, while existing incremental methods underperform on fraud detection and retraining is costly.
Approach
Combines a contrastive-style pre-trained GraphSAGE encoder on positive/negative claim graphs with a GAT detector, and updates parameters online using a momentum-based Temporal MAS importance penalty that needs no stored historical data.
Key results
- On the medical fraud dataset, POCL reaches 80.34 AUC and 63.53 F1, close to RetrainGNN (80.58 AUC, 63.77 F1) but with 3.62 min total time versus 97.4 min for RetrainGNN.
- POCL outperforms the other online and fraud-detection baselines by roughly 1%-4% in average monthly accuracy on the medical, Amazon, and YelpChi datasets.
- Ablations show removing either the contrastive pre-training or Temporal MAS lowers long-term accuracy, and the combined model reduces the decline observed in each part.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use POCL instead of RetrainGNN | Medical fraud detection · Medical fraud dataset | AUC: 80.58 -> 80.34 | No clear effect |
| Use POCL instead of RetrainGNN | Medical fraud detection · Medical fraud dataset | F1: 63.77 -> 63.53 | No clear effect |
| Use POCL instead of RetrainGNN | Medical fraud detection · Medical fraud dataset | Time (minutes): 97.4 -> 3.62 | Helped |
| Use POCL instead of RetrainGNN | Medical fraud detection · Medical fraud dataset | Space (GB): 11.4 -> 2.35 | Helped |
| Remove Temporal MAS and directly update parameters (ablation) | POCL · Medical fraud dataset | Accuracy | Hurt |
| Add Temporal MAS online updates (ablation) | POCL without Temporal MAS · Medical fraud dataset | Accuracy | Helped |
| Remove pre-trained contrastive learning and use GraphSAGE (ablation) | POCL · Medical fraud dataset | Accuracy | Hurt |
| Add pre-trained contrastive learning (ablation) | POCL without pre-trained contrastive learning · Medical fraud dataset | Accuracy | Helped |
Surprising findings
- Its pre-trained and MAS-regularized model performs almost as well as full retraining while using about 1/27 of the training time.
Limitations
- Results compare against only a limited set of baselines and no LLM or modern fraud-detection methods.
- The penalty uses gradients as importance weights with a hand-set momentum and penalty weight, with limited sensitivity analysis.
- Online updates are only evaluated over simulated or single-year settings; the Amazon and YelpChi evaluation uses simulated timestamps.
Experimental setup
Datasets: medical fraud dataset (Ma et al. 2023) Amazon YelpChi
Benchmarks: medical fraud dataset Amazon YelpChi
Metrics: accuracy AUC F1 training time memory consumption
Baselines: OnlineGNN OfflineGNN RetrainGNN ERGNN FGN ContinuesGNN CAREGNN PCGNN
Builds on: Memory Aware Synapses GraphSAGE GAT Deep Graph Infomax
Closest papers by meaning
- Online Fraud Detection via Test-Time Retrieval-Based Representation Enrichment (AAAI 2025 · similarity 0.82)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.82)
- Scale-Aware Contrastive Reverse Distillation for Unsupervised Medical Anomaly Detection (ICLR 2025 · similarity 0.81)
- Barely Supervised Learning for Graph-Based Fraud Detection (AAAI 2024 · similarity 0.80)
- DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.80)
- Leveraging Catastrophic Forgetting to Develop Safe Diffusion Models against Malicious Finetuning (NeurIPS 2024 · similarity 0.79)
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.