Explainable Scene Understanding with Qualitative Representations and Graph Neural Networks

Graph-based reasoning improves explainability in autonomous vehicles: The paper demonstrates that using Graph Neural Networks to analyze entire traffic scenes—rather than just pairs of objects—enables autonomous driving systems to better explain why other vehicles or pedestrians are behaving in certain ways, which is critical for safe decision-making.

Explainability requires understanding context, not just individual relationships: Previous approaches only examined single relationships between two objects at a time. This work shows that considering the "complete spatial-temporal relationships between all objects in the scene" is necessary for truly understanding what's happening in traffic, suggesting that responsible AI systems need broader contextual awareness.

Handling real-world data imbalances is essential for practical deployment: The model "effectively handles the inherent class imbalance in relevant object identification tasks," addressing a common real-world challenge where some objects in a scene are far more important than others—a practical consideration for building trustworthy autonomous systems.

Human annotation provides the ground truth for explainability: The research relies on "human-annotated relevance labels" from the DriveLM dataset, highlighting that developing explainable AI systems requires human experts to define what explanations should look like, rather than letting algorithms decide independently.

Combining symbolic and deep learning approaches strengthens explainability: By integrating Qualitative Explainable Graphs (a symbolic, human-interpretable method) with Graph Neural Networks (a deep learning approach), the work suggests that hybrid methods may offer better explainability than either approach alone.