Abstract
Detecting medical insurance fraud is a dynamic challenge as fraudsters continuously adapt their tactics. Traditional static models fail to keep pace with these shifts, while incremental learning often suffers from catastrophic forgetting. This paper introduces POCL (Pre-trained Online Contrastive Learning), a framework that leverages contrastive pre-training and a momentum-based parameter update strategy to achieve real-time fraud detection. POCL effectively balances the learning of new fraud patterns with the retention of historical knowledge, offering a significant improvement in accuracy and computational efficiency over state-of-the-art methods.
Introduction
Medical insurance fraud results in substantial financial losses for healthcare systems globally. The primary difficulty in detection lies in the "cat-and-mouse" nature of fraud, where perpetrators evolve their methods to bypass existing security measures. Static detection models, trained on fixed datasets, quickly become obsolete. While online learning offers a potential solution, it is frequently hampered by catastrophic forgetting—the tendency of neural networks to overwrite previously learned information when trained on new data. Conversely, retraining models from scratch is often infeasible due to the sheer volume of insurance data.
Methodology
The POCL framework addresses these challenges through a three-stage pipeline designed for stability and plasticity:
- Contrastive Pre-training: The model is first trained on historical data to map medical claims into a feature space where fraudulent and normal nodes are distinct. This stage establishes a robust foundation for identifying risk by clustering similar labels and separating dissimilar ones.
- Task-Specific Learning: The pre-trained features are integrated into a Graph Attention Network (GAT). A specialized loss function is employed to ensure the GAT learns classification tasks without losing the structural insights gained during pre-training.
- Online Learning with Temporal MAS: To handle evolving data, POCL utilizes Temporal Memory Aware Synapses (MAS). This mechanism calculates parameter importance for the current task and uses a momentum-based update to gradually phase out the importance of older weights. This allows the model to adapt to current trends while maintaining relevant historical knowledge without storing old data.
Experimental Evaluation
The researchers evaluated POCL using a large-scale real-world medical insurance dataset (~550,000 claims) and standard fraud benchmarks (Amazon and YelpChi). For the benchmarks, timestamps were simulated to evaluate performance in a streaming data context.
Results and Discussion
POCL demonstrated superior performance across multiple metrics compared to state-of-the-art (SOTA) models like CAREGNN and PCGNN:
- Accuracy and Stability: POCL achieved a 1% to 4% higher average monthly accuracy than other online models. Its performance closely matched that of a model retrained monthly (RetrainGNN), but with significantly lower overhead.
- Resilience to Performance Decay: As fraud patterns shifted over a 12-month period, POCL maintained stable accuracy, whereas competing online methods experienced sharp declines.
- Efficiency: POCL proved to be approximately 27 times faster than full retraining (3.6 minutes vs. 97.4 minutes) and required 5 times less memory (2.35 GB vs. 11.4 GB).
The success of POCL is attributed to the use of contrastive learning as an "anchor" for evolving characteristics and the momentum-based update strategy, which effectively manages the trade-off between learning new information and preserving old knowledge. Furthermore, by avoiding the storage of historical data for updates, POCL enhances privacy and reduces storage requirements.
Conclusion
POCL offers a scalable and efficient solution for real-time insurance fraud detection. While currently focused on node-level classification, future research could extend this approach to detect fraud within complex subgraphs or relational edges. Additionally, the effectiveness of the online phase remains dependent on the quality of the initial pre-training dataset.
Summary updated