Nomograms for Visualizing Linear Support Vector Machines

Support vector machines are often considered to be black box learning algorithms. We show that for linear kernels it is possible to open this box and visually depict the content of the SVM classifier in high-dimensional space in the interactive format of a nomogram. We provide a crosscalibration method for obtaining probabilistic predictions from any SVM classifier, which control for the generalization error. If we employ logistic regression for calibration, the effect of each attribute can be represented on the log odds ratio scale. We also describe an approach to capturing nonlinear effects of continuous attributes with an ordinary linear kernel, and adapt the nomogram so that these nonlinear effects can be graphically rendered.