Identification of High-Risk Patterns in Automobile Insurance Claims Using Deep Representation Learning and Evolutionary Clustering

Abstract

Detecting fraud in automobile insurance claims is challenging because reliable labels are scarce, the classes are highly imbalanced, and fraudulent behavior is often concealed. Supervised methods depend on accurately labeled training data, whereas unsupervised methods can identify unusual patterns but may offer limited support for prioritizing high-risk claims. This study proposes a hybrid framework that combines deep representation learning and evolutionary clustering to identify and prioritize potentially high-risk automobile insurance claims. The framework was evaluated using a real-world dataset of 179,514 automobile insurance claims. A deep autoencoder learned compact latent representations, and a genetic algorithm optimized initial cluster centroids in the latent space. K-means was then used for final cluster assignment. The main evaluation configuration used a 16-dimensional latent space, 40 training epochs, the Adam optimizer, and five clusters. A composite risk score combined reconstruction error, distance from the assigned cluster centroid, and insurance-specific profile indicators. Evaluation used a limited set of verified labels and synthetically injected fraudulent cases, with Precision@k, Recall@k, and the area under the precision–recall curve (PR-AUC) as the main performance measures. The proposed method produced more compact and better-separated clusters than the baseline methods. The silhouette score increased from 0.124 for K-means in the raw feature space to 0.3953 for evolutionary clustering in the autoencoder latent space. Precision@100 reached 0.68, compared with 0.18 for the K-means baseline, and PR-AUC increased from 0.142 to 0.562. The model identified groups with unusual financial and temporal characteristics, supporting the prioritization of suspicious claims for further investigation. Combining deep representation learning with evolutionary clustering offers a practical decision-support approach for ranking automobile insurance claims and directing audit resources toward potentially high-risk cases. The limited availability of verified fraud labels and the use of synthetic evaluation cases constrain conclusions about real-world fraud detection performance. The framework should therefore support expert review rather than determine whether a claim is fraudulent.