Diffusion Integrated Gradients: Controllable Path Generation for Flexible Feature Attribution

The paper introduces Diffusion Integrated Gradients (DiffIG), a novel method for generating attribution paths in model explanations by reformulating path generation as a conditional generative modeling problem using a diffusion model. DiffIG learns a distribution over paths via a Stick-Breaking Process and incorporates user guidance during sampling, leading to improved quality of feature attributions compared to traditional methods. This approach enhances Explainable AI (XAI) by providing practitioners with a flexible and controllable means of generating explanations, potentially leading to more accurate insights into model behavior.