SVAP: Shapley Value Guided Attribution Prior for Neural Network-Based Autonomous Driving

Deep neural networks (DNNs) have demonstrated remarkable performance in safety-critical applications such as autonomous driving, yet their reliability remains constrained by inherent dependencies on confounding factors. Recent advances in feature attribution methods have shown potential for integrating prior knowledge to enhance model interpretability and performance. However, current attribution priors frequently fail to satisfy fundamental principles, including efficiency, dummy player, symmetry, and linearity axioms, while neglecting feature interactions. Moreover, existing methods lack validation in complex real-world driving scenarios where multiple dynamic agents interact under uncertain environmental conditions. We introduce a Shapley value-based attribution prior (SVAP) approach to address these shortcomings. Originating from cooperative game theory, the Shapley value offers a set of axioms that provide theoretical guarantees, making it an ideal tool for improving attribution priors. Within the SVAP framework, we incorporate importance and smoothness priors into a deep reinforcement learning-based autonomous driving decision model and an image classification task focusing on traffic participants. The empirical results demonstrate the effectiveness and reliability of the proposed method. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/limeng-1234/SVAP</uri>.