Probabilistic Path Integration with Mixture of Baseline Distributions

Path integration methods generate attributions by integrating along a trajectory from a baseline to the input. These techniques have demonstrated considerable effectiveness in the field of explainability research. While multiple types of baselines for the path integration process have been explored in the literature, there is no consensus on the ultimate one. This work examines the performance of different baseline distributions on explainability metrics and proposes a probabilistic path integration approach where the baseline distribution is modeled as a mixture of distributions, learned for each combination of model architecture and explanation metric. Extensive evaluations on various model architectures show that our method outperforms state-of-the-art explanation methods across multiple metrics.