ADAPTIVE IMBALANCE LEARNING FOR ACCURATE HEALTH INSURANCE FRAUD DETECTION USING META-REINFORCEMENT LEARNING

Abstract

A meta-reinforcement learning system that can react to high-class imbalances and changing fraud patterns would be perfect for health insurance fraud detection. The suggested method detects minority (fraud) classes by automatically determining sampling, weighting, and decision-making procedures. Meta-learning lets the system adapt to new fraud patterns without retraining. Reinforcement learning improves policy using sparse and delayed reward sets using real-world claims data. A hybrid deep representation module tracks claimants, suppliers, and recipients' complicated temporal and interpersonal links. The design reduces fraud detection false negatives with cost-sensitive learning. Experimental results on real-world and standard health insurance datasets show increased F1-score, recall, and precision over static classifiers and conventional resampling. This makes hostile adaptation and notion drift harder for fraudsters. Scalability comes via gradual improvements and online learning. The model's clear decision indicators help inspectors prioritize inquiries.