Dynamic Risk Optimization in Enterprise Decisions Via Dual-Attention Temporal Graph Reinforcement Learning

Abstract

Modern business settings are complex and risk-sensitive, requiring sophisticated and adaptable solutions for informed organizational decision-making. Existing solutions use static or rule-based models that cannot dynamically analyze real-time decision risks. This research introduces DynaRisk-OptNet, a deep learning system for enterprise decision risk optimization that combines Hierarchical Dual-Attention Temporal Graph Reinforcement Network (HDAT-GRN) and Soft Actor-Critic (SAC) reinforcement learning. The model dynamically captures temporal dependencies, cross-feature interactions, and structural risk propagation. Dual attention weights and gradient-based saliency improve interpretability. Results on a real-world enterprise risk dataset showed that the system outperformed recent transformer-based benchmarks with a TRPE of 0.93, an AASI of 3.8, and an FAFS of 0.89 for feature attribution fidelity. To achieve scalability and high inference speed (17.6 ms/sample), the implementation made use of PyTorch and DGL. These findings confirm that the model is both practically applicable and easily explicable, making it an excellent choice for fast-paced, high-stakes business settings. As a result, DynaRisk-OptNet offers a robust and intelligent framework for risk-aware organizational decision optimization.