Generating Explanations from Support Vector Machines for Psychological Classifications

Abstract

An explanation capability is crucial in security-sensitive domains, such as medical applications. Although support vector machines (SVMs) have shown superior performance in a range of classification and regression tasks, SVMs, like artificial neural networks (ANNs), lack an explanatory capability. There is a significant literature on obtaining human-comprehensible rules from SVMs and ANNs in order to explain “how” a decision was made or “why” a certain result was achieved. This chapter proposes a novel approach to SVM classifiers. The experiments described in this chapter involve a first attempt to generate textual and visual explanations for classification results using multimedia content of various type: poems expressing positive or negative emotion, autism descriptions, and facial expressions, including those with medical relevance (facial palsy). Learned model parameters are analyzed to select important features, and filtering is applied to select feature subsets of explanatory value. The explanation components are used to generate textual summaries of classification results. We show that the explanations are consistent and that the accuracy of SVM models is bounded by the accuracy of explanation components. The results show that the generated explanations display a high level of fidelity and can generate textual summaries with an error rate of less than 35 %.