Local Surrogates vs Global Surrogates: visual comparison of decision boundaries

This work explores new methods for comparing local and global surrogates through decision region visualization. We present an implementation for approximating and visualizing decision regions of text classification models. The implementation combines text perturbation, dimensionality reduction and Voronoi diagrams. The proposed method provides more insights about decision boundary landscape complexity compared to basic numerical scores such as accuracy or the R-squared metric.