An Open-Source Software Library for Explainable Support Vector Machine Classification

Abstract

Since computer predictions can be used for supporting high-impact decision-making processes, finding mechanisms that favor the interpretability of such predictions has grown into a challenging research topic. One of those mechanisms proposes the contextualization of support vector machine (SVM) predictions within a classification process by means of the most influential support vectors (MISVs). An open-source software library by which a researcher or practitioner can obtain such contextualized SVM predictions is proposed and presented in this paper. The proposed library enables the creation of coherent explanations for SVM predictions about the class(es) of an object to be built from contextualized evaluations of the membership (and nonmembership) of that object in different classes. Examples illustrating how to use the library for designing and building explanation interfaces are presented.