SPINE: VAE‐driven Counterfactuals for Decision Boundary Maps

As Deep Learning models become increasingly complex, Explainable AI becomes essential for deploying machine learning classifiers. Decision Boundary Mapping (DBM) is a technique for visualizing a classifier's global decision boundary. Despite their relative success, current DBM methods rely on global inverse multidimensional projections that map 2D points back to the input space. The resulting visualizations depend heavily on the chosen projection, distort elements of the structure and distribution of the input data, and contain limited classifier‐specific information, often leading to inaccurate boundaries. We propose Sampling‐based Precise Neighbourhood Estimation (SPINE), a novel DBM approach for differentiable classifiers that is projection‐agnostic and does not depend on inverse projections. SPINE generates additional high‐dimensional points near the decision boundary using counterfactuals and variational autoencoders, then applies local interpolation to build visual representations. Our results show that this strategy better captures the structure and distribution of the data and the classifier's predictions in boundary regions than the state‐of‐the‐art in DBMs, yielding higher‐resolution and more accurate representations of classifier behaviour. Beyond standard DBM representations, SPINE enables additional exploratory analyses, such as counterfactual feature changes and path visualizations, supporting richer insights into model decisions.