Efficient Handling of Data Imbalance in Health Insurance Fraud Detection Using Meta-Reinforcement Learning

Abstract

Data imbalance is one of the major challenges in health insurance fraud detection where the distribution of classes within the dataset is significantly skewed, leading statistical models to be biased toward the dominant class. The algorithmic approaches to handling imbalance involve modification to the loss functions to sensitize them to the minor class. Our research focuses on the efficiency of meta-models in addressing imbalance and introduces Meta-reinforcement learning (Meta-RL) as a novel solution for learning under imbalance. Meta-RL leverages meta-learning principles to learn the characteristics of fraudulent instances dynamically. This adaptability comes from its ability to learn shared representations and optimal task-specific strategies using limited samples. By using task distribution and reward shaping, our experiments on Meta-RL algorithms, RL2, and VariBAD achieve superior sample efficiency and adaptability for varying degrees of imbalance. The efficiency of the models is measured using imbalance-safe metrics like Geometric mean, Harmonic mean, and Mathew's Correlation Coefficient (MCC), and metrics like Cohen's Kappa score are used to gauge the consistency of the results. This research is the first to apply Meta-RL to the problem of data imbalance in fraud detection, contributing to a generalizable and efficient framework for imbalanced learning. The findings of this research show that Meta-RL algorithms can be effectively tuned to handle data imbalance without modification to their objective functions and hence, can be considered an appropriate option for health insurance fraud detection solutions.