Title: Attribution via Distributional Paths for Information Revelation

Abstract:

Feature attribution techniques elucidate model predictions by allocating significance scores to various input components. Path-based approaches, including Integrated Gradients, are particularly attractive due to their adherence to the completeness property, ensuring that attributions collectively equal the difference in model output between a baseline state and the actual input. However, conventional path methods typically define this trajectory within the input space, interpreting the model through pointwise perturbations along a specific route. This approach integrates the model’s raw responses at each point along the path, offering no mechanism to control the resolution of feature queries. Consequently, the initial segment of the trajectory, which lies close to the baseline, holds equal explanatory weight as the final input itself.

In this work, we elevate path attribution from the input space to a space of structured probe distributions centered on the example of interest. We term this new approach Reveal-IG. Instead of moving through raw input values, Reveal-IG gradually discloses information regarding the input and assigns importance to changes in the model’s expected output along this distributional trajectory. This framework preserves completeness relative to the expected model response while inherently supporting multiscale image probes and feature-wise uncertainty in tabular datasets. Synthetic evaluations demonstrate that Reveal-IG mitigates path artifacts commonly found in input-space methods. Furthermore, in tasks involving ImageNet classification and tabular regression, the method generates stable, signed attributions, achieving top performance on metrics reliant on attribution signs while maintaining competitiveness across other benchmarks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC