Understanding Support Vector Machine Classifications via a Recommender System-Like Approach

Support vector machines are a valuable tool for making classifications, but their black-box nature means that they lack the natural explanatory value that many other classifiers possess. Alternatively, many popular websites have shown recent success in explaining recommendations based on behavior of other users. Inspired by these ideas, we suggest two novel methods for providing insight into local classifications produced by a support vector machine. In the first, we report the support vectors most influential in the final classification for a particular test point. In the second, we determine which features of that test point would need to be altered (and by how much) in order to be placed on the separating surface between the two classifications. We call the latter technique “border classification.” In addition to introducing these explanatory techniques, we also present a free-for-download software tool that enables users to visualize these insights graphically.