Semantic Pareto-DQN: Multi-Objective RL Framework for Financial Anomaly Detection

The authors propose Semantic Pareto-DQN, a multi-objective reinforcement learning framework designed to address extreme class imbalance and "fraud collapse" in financial anomaly detection. The system synthesizes heterogeneous transaction features into natural-language narratives encoded by large language models to create robust state representations.

• Optimizes a vectorial reward that decouples financial efficacy, operational friction, and semantic discovery.
• Maps the continuous Pareto frontier to dynamically navigate costs of missed anomalies versus false positives.
• Achieves superior minority-class recall on E-Commerce fraud and UCI Credit datasets compared to scalarized baselines.

This approach provides an alternative to data resampling by balancing anomaly interdiction with customer friction without distorting the underlying data distribution.