Insurance Fraud Detection: A Perspective with S-LIME Explanability
Gian Reinfred Athevan, Felix Indra Kurniadi, Thomas Amarta Gunawisesa, Anderies, Imanuel Tio · 2024
Healthcare insurance has a frequent problem: fraudulent claims. As a result, it is not surprising that the application of machine learning techniques and analytics to investigate claim behavior is a well-established issue in the literature. Despite AI's effectiveness in detecting fraudulent actions, especially in the insurance industry, the black box aspect of AI models raises serious concerns. These models frequently lack clarity, making it impossible to determine the significance of each feature in the final conclusion. This opacity serves as both a barrier and an opportunity for fraudsters, who are constantly adapting their techniques to avoid discovery. Transparency is especially important in the context of the General Data Protection Regulation (GDPR), which requires explanations for choices made using automated processing. This paper argues for the need of understanding what the model learns and how to improve it, emphasizing that customers have the right to know the outcome of machine predictions as well as the logic behind them. The study seeks to investigate answers to these difficulties, opening the way for more visible, explainable, and accountable usage of AI in insurance fraud detection. This study investigated the use of S-LIME, a model-agnostic explanation technique, to identify influential features within various models. Even with a substantial reduction in features, an increase in model accuracy by 1 % can still be achieved in XGBoost, or the accuracy can remain unchanged as observed in Random Forest. This analysis revealed that S- LIME successfully pinpoints features with a greater impact on prediction.