Feature attribution methods that respect dependencies between input features often struggle with high-dimensional data. According to a new arXiv preprint, these existing approaches simply do not scale, limiting their practical use in modern machine learning settings.
The authors introduce causal integrated gradients (CIG) and causal expected gradients (CEG), which extend the well-known integrated gradients framework to operate along causal paths. This design aims to preserve the benefits of dependency-aware attribution while achieving scalability to high dimensions.
The abstract does not yet provide experimental details or comparisons, but the proposal directly targets a known bottleneck in explainability research. If the scalability claims hold, CIG and CEG could offer a practical tool for interpreting complex, high-dimensional models where causal structure is relevant.