Title: ShapDBM: Investigating Decision Boundary Maps within Shapley Space

Abstract:

Decision Boundary Maps (DBMs) serve as a potent method for visualizing classification boundaries in machine learning. However, the fidelity of these maps is heavily contingent upon the dimensionality reduction (DR) method employed and the high-dimensional space from which the data points are drawn. In the context of complex machine learning datasets, dimensionality reduction often results in mixed classes, producing DBMs that are either difficult to interpret or potentially deceptive. To address this, we introduce a novel approach that generates DBMs by first transforming the data space into Shapley space before applying dimensionality reduction. When contrasted with DBMs derived directly from the raw data, our proposed maps demonstrate comparable or superior quality metrics. Furthermore, they offer decision zones that are visibly more compact and intuitive to navigate, while also aligning more closely with the model’s measured performance.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC