Interactive Evolutionary Optimization of Visual Explainable AI through Gestalt Principles with Human Feedback

In the context of eXplainable Artificial Intelligence (XAI), describing the rationale behind a Machine Learning (ML) model’s decisions is crucial to user understanding and acceptability. One way to facilitate this is through visualization of the model’s behavior and its resulting decisions. In this paper, we present a novel Visual-based XAI (vXAI) framework that co-designs the model with its graphical representation. Specifically, we consider a training phase in which a classification task is solved separately on two training datasets. For each dataset, we select the best among different trained classifiers, on which we then perform feature selection. Next, visualizations of the classifiers’ decision maps are presented to users, who rank them based on clarity of explanation. Interactive Constrained MAP-Elites is then used to optimize an aggregate score that weighs five visual metrics inspired by Gestalt cognitive principles, which all correlate with a clearer visual understanding of th...