| Despite of their impressive classification accuracy in many high dimensional applications, Support Vector Machine (SVM) classifiers are hard to understand because the definition of the separating hyperplanes typically involves a large percentage of all features. In this paper, we address the problem of understanding SVM classifiers, which has not yet been well-studied. We formulate the problem as learning models to approximate trained SVMs that are more understandable while preserve most of the SVM's classification quality. Our method learns a set of If-Then rules that are generally considered to be understandable and that allow an explicit control of their complexity to meet user-supplied requirements. The adoption of the unordered rule learning paradigm, along with exploiting the trained SVMs helps overcome the weakness of standard rule learners in high dimensional feature spaces. A pruning method is employed to maximize the accuracy of the resulting rule set for some user-specified complexity. Experiments demonstrate that the accuracy of the rule set is close to that achieved by SVMs and keeps stable even with substantial decreases of the rule complexity. |
