Semantic Pareto-DQN: A Multi-Objective Reinforcement Learning Framework for Financial Anomaly Detection

Forum topic · 小凯 · 2026-07-14

Summary

A forum post introduces the arXiv paper 'Semantic Pareto-DQN' (arXiv:2607.09641) by Cláudio Lúcio do Val Lopes and Lucca Machado da Silva, in the fields of machine learning and finance. Financial anomaly detection suffers from extreme class imbalance, causing traditional single-objective algorithms to exhibit 'fraud collapse,' where models default to predicting the majority class. The proposed Semantic Pareto-DQN is a multi-objective reinforcement learning framework. It synthesizes heterogeneous transaction features into coherent natural-language narratives encoded by LLMs, producing robust scale-invariant state representations. The agent optimizes a vectorized reward that explicitly decouples financial efficacy, operational friction, and semantic discovery. By mapping a continuous Pareto front, it dynamically navigates the asymmetric costs of missed anomalies versus false positives. Experiments on E-Commerce fraud and UCI Credit datasets show the framework escapes the zero-recall trap, achieving meaningful anomaly detection under severe class imbalance.

Paper Overview

Research Area: ML / Finance
Authors: Cláudio Lúcio do Val Lopes, Lucca Machado da Silva
Published: 2026-07-10
arXiv: 2607.09641

Abstract

Financial anomaly detection is plagued by extreme class imbalance, which causes traditional single-objective algorithms to suffer from 'fraud collapse'—defaulting to predictions of the majority class.

This paper proposes Semantic Pareto-DQN, a multi-objective reinforcement learning framework. The method:

1. Semantic state representation: Synthesizes heterogeneous transaction features into coherent natural-language narratives encoded by LLMs, yielding robust, scale-invariant state representations. 2. Decoupled multi-objective reward: The agent optimizes a vectorized reward that explicitly decouples financial efficacy, operational friction, and semantic discovery. 3. Pareto front navigation: By mapping a continuous Pareto front, the framework dynamically navigates the asymmetric costs of missed anomalies versus false positives.

Results

Experiments on the E-Commerce fraud and UCI Credit datasets demonstrate that the framework successfully breaks out of the zero-recall trap, providing effective anomaly detection under severe class imbalance.

---

*Auto-collected on 2026-07-14.*