MOUNTAINEER: Topology-Driven Visual Analytics for Comparing Local Explanations

• DOI: 10.48550/arxiv.2406.15613
• Publication Date: 2024-06-21
• Authors: Parikshit Solunke; Vitória Guardieiro; João Rulff; Peter Xenopoulos; Gromit Yeuk-Yin Chan; Brian Barr; Luís Gustavo Nonato; Claudio Silva
• Keywords: Mountaineering

With the increasing use of black-box Machine Learning (ML) techniques in critical applications, there is a growing demand for methods that can provide transparency and accountability model predictions. As result, large number local explainability models have been developed popularized. However, machine learning explanations are still hard to evaluate compare due high dimensionality, heterogeneous representations, varying scales, stochastic nature some these methods. Topological Data Analysis (TDA) be an effective method this domain since it used transform attributions into uniform graph providing common ground comparison across different explanation We present novel topology-driven visual analytics tool, Mountaineer, allows ML practitioners interactively analyze representations by linking topological graphs back original data distribution, predictions, feature attributions. Mountaineer facilitates rapid iterative exploration explanations, enabling experts gain deeper insights techniques, understand underlying distributions, thus reach well-founded conclusions about behavior. Furthermore, we demonstrate utility through two case studies using real-world data. In first, show how enabled us discern regions causes disagreements between explanations. second, tool themselves. Finally, conducted interviews with three industry help our work.