Abstract: Autonomous vehicles operating at urban intersections face substantial behavioral decision-making uncertainty due to complex multi-vehicle interactions and incomplete environmental information. To address this challenge, this paper proposes a deep reinforcement learning-based behavior decision method for autonomous vehicles, termed Graph Attention and GRU enhanced Deep Deterministic Policy Gradient (GAG-DDPG) . First, both epistemic and aleatoric uncertainties in autonomous driving scenarios are modeled and quantified to facilitate traffic risk assessment. To effectively capture the dynamic interactions among surrounding vehicles and their temporal dependencies, a Graph Attention Network (GAT) and a Gated Recurrent Unit (GRU) are integrated into the policy network. Specifically, the GAT employs a multi-head attention mechanism to model inter-vehicle interaction relationships and adaptively focus on critical surrounding vehicles, while the GRU captures the temporal evolution of traffic states. This design enables the construction of a spatiotemporal interaction-aware state representation, providing a reliable feature basis for uncertainty-aware risk modeling. Furthermore, Conditional Value at Risk (CVaR) is incorporated to develop a risk-sensitive decision-making mechanism, allowing the agent to select more robust behavioral actions under uncertain environments and thereby achieve a balance between safety and exploration. Experimental results demonstrate that the proposed GAG-DDPG method significantly improves both the success rate and driving smoothness of autonomous vehicles in intersection left-turn scenarios, thereby enhancing decision-making performance in complex and uncertain traffic environments.