Semantic Pareto-DQN: A Multi-Objective Reinforcement Learning Framework for Financial Anomaly Detection
Semantic Pareto-DQN is a multi-objective reinforcement learning framework for financial anomaly detection, designed to address the 'fraud collapse' problem caused by extreme class imbalance without using distorted data resampling. The framework was published on arXiv.
Semantic Pareto-DQN is a multi-objective reinforcement learning framework for financial anomaly detection that balances anomaly interception and customer friction by optimizing multiple objectives, avoiding the 'fraud collapse' problem that occurs in traditional single-objective algorithms due to class imbalance.
The framework handles class imbalance through multi-objective optimization, potentially incorporating semantic information to guide Pareto frontier search. Next steps could include evaluating its performance on datasets with higher imbalance ratios and comparing it with other multi-objective RL methods.
The financial anomaly detection field has long been plagued by class imbalance. This framework offers an alternative without resampling. If effective, it could drive a shift from single-objective to multi-objective decision-making in risk control systems.
If successful, it could reduce fraud losses while minimizing false positives that disrupt users, improving risk control efficiency and customer experience for financial institutions.
Future work may involve deployment testing on real financial transaction data and integration with other anomaly detection methods such as GANs and autoencoders.