An Explainable AI Multi-Agent Recommender System for Financial Document Access Control
提出一个多智能体推荐系统，结合FinBERT、BERT和GPT模型对金融文档进行敏感度分类（公开、内部、机密、限制），由GPT-5.1协调生成可解释的访问控制建议，准确率达83.71%，且一致决策时准确率更高。
Abstract Financial institutions require robust document access control mechanisms that balance security with transparency and explainability. Traditional classification systems often operate as black boxes, failing to provide justifications for access-control decisions. This work presents a novel explainable AI multi-agent recommender system for financial document sensitivity classification that addresses critical ethical concerns in AI-powered decision-making. We fine-tuned three state-of-the-art models—FinBERT, BERT-base-uncased, and GPT-4.1-mini—on a custom-labeled Financial PhraseBank dataset with four sensitivity levels: Public, Internal, Confidential, and Restricted. These fine-tuned models serve as specialized AI agents within a multi-agent architecture orchestrated by GPT-5.1, a large reasoning model operating in zero-shot mode. The orchestrator synthesizes agent predictions and generates natural language recommendations that justify classification decisions. Our agentic AI multi-agent recommender system achieves 83.71% overall accuracy, comparable to individual models (82.80%-84.93%), while providing interpretable explanations for each decision. Critically, agent agreement analysis reveals that unanimous decisions (3/3 agents agree, 78.8% of cases) achieve 92.28% accuracy—significantly outperforming any individual model—validating the collaborative decision-making approach. The system demonstrates that multi-agent architectures can provide both high-confidence predictions and natural language explainability, creating transparent, accountable AI systems for financial document access control. All code and methodologies are released as open-source on our GitHub (Applied-AI-Research-Lab 2025) to support reproducibility and further research in explainable AI for finance.