Attribution explanations for decision-making in deep lane-change models

Abstract

Deep learning models are attracting considerable attention for their potential to enable intelligent lane-change behaviors. To ensure the reliability of these models, it is essential to understand their decision-making processes using attribution methods. However, existing attribution techniques, which are predominantly developed for visual tasks, often struggle to deliver accurate and interpretable explanations when applied to complex lane-change models. We identify the essential cause of this problem as the high variability in the distribution of connected perceptual information that serves as input to lane-change models. To address this, we propose a novel path aggregation attribution method, where paths describe the transition of a feature from absence to presence, expressing its relative contribution. Our method leverages an exponential family to introduce probabilistic paths and calculate attribution expectations, effectively traversing the input feature distribution space to provide a more comprehensive representation of feature transitions. Additionally, we introduce a distribution-informed counterfactual reference to define starting points of the paths, enabling the flexible generation of traffic scenarios with feature absence. Extensive experiments on three lane-change models show that our method consistently outperforms state-of-the-art attribution methods. Specifically, we achieve higher performance on four widely used quantitative metrics, i.e., sensitivity-n, accuracy information curve, softmax information curve, and most-relevant-first, demonstrating superior reliability and interpretability.