Learning Dual-Path Soft Decision Trees for Vision Prototype XAI

Following the principle of “this part supports that decision,” prototype-based models achieve visual task reasoning by matching parts to reference patterns. While the explanations are intuitive, they are not locally faithful and fail to quantify the contribution of individual features. In this work, we introduce an attention mechanism that pinpoints discriminative regions and exposes the decision logic for explanation generation. We proposeSpaViTree(Spatial Visual Attention Decision Tree), a framework combining soft decision trees and dual-path feature processing for spatially-grounded interpretability. At each decision node, one stream captures fine-grained and localized visual evidence through attention-guided features, and the other retains global semantic context via simple feature transformations. A learned gating mechanism, informed by region selection signals, adaptively combines the two streams to guide node-level decisions. Experiments demonstrate competitive classification accuracy alongside superior interpretability, with visualizations highlighting advantages over existing prototype-based models. Quantitative validation using insertion-deletion scores confirms heatmap faithfulness, and rule statistics address the requirements of tree structure analysis.