Stability Analysis of Supervised Decision Boundary Maps

Understanding how a machine learning classifier works is an important task in machine learning engineering. However, doing this is for any classifier in general difficult. We propose to leverage visualization methods for this task. For this, we extend a recent technique called Decision Boundary Map (DBM) which graphically depicts how a classifier partitions its input data space into decision zones separated by decision boundaries. We use a supervised, GPU-accelerated technique that computes bidirectional mappings between the data and projection spaces to solve several shortcomings of DBM, such as accuracy and speed. We present several experiments that show that SDBM generates results which are easier to interpret, far less prone to noise, and compute significantly faster than DBM, while maintaining the genericity and ease of use of DBM for any type of single-output classifier. We also show, in addition to earlier work, that SDBM is stable with respect to various types and amounts of changes of the training set used to construct the visualized classifiers. This property was, to our knowledge, not investigated for any comparable method for visualizing classifier decision maps, and is essential for the deployment of such visualization methods in analyzing real-world classification models.