Computing fast and accurate maps for explaining classification models

Image representations of the behavior of trained machine learning classification models can help machine learning engineers examine various aspects of a model such as how it partitions its data space into decision zones separated by decision boundaries; how training samples support the decision in various parts of the data space; and how close training data is to decision boundaries. Yet, for an image of n × n pixels, all current methods that create such images have a computational complexity of O ( n 2 ) which precludes their use in interactive visual analytics scenarios. We present a set of techniques for the fast computation of such image-based classifier representations. Compared to earlier work in this area, we accelerate both so-called decision maps, that compute categorical labels, and classifier maps, that compute real-valued quantities, in O ( ( log n ) 2 ) time. Practically, our method has a speed-up of about one order of magnitude and yields results very similar to the ground-truth maps; has no free parameters; is model agnostic; and is simple to implement. We demonstrate our method on several combinations of maps, datasets, and classification models. • We present FastDBM, a fast method for computing 2D maps (images) that match ground-truth results while avoiding costly per-pixel evaluations. • Our method generalizes earlier approaches from categorical to real-valued functions, supporting any (smooth) function. • We evaluate FastDBM with 3 projection techniques, 4 inverse projection techniques, 4 datasets, and 6 classifiers, showing consistent speedups and high-quality results.