Toward Actionable Support Vector Machines: A Ranking-based Approach
During the last decade, Support Vector Machines (SVMs) have attracted a great deal of attention and achieved huge success mainly as powerful classifiers. However, one of the main drawbacks of this learning method is the lack of intelligibility of the results. SVMs are "black box" systems that do not provide insights on the reasons of a classification or explanations - the results produced must be taken on faith. We are concerned about the problem of intelligibility because from our practical experience, domain experts strongly prefer Machine Learning with explanations rather than a black box even if the black box system achieves a high predictive performance. In that context, we have developed a new approach to provide explanations and make SVMs results more actionable. The underlying idea is to produce explanations by applying symbolic Machine Learning models to SVM-produced ranking results. More precisely, we are contrasting SVM results from the top and bottom of rankings to detect the main discriminative properties between classes which can be quite useful for the practitioner to direct actions and understand the system. We applied our approach on several datasets. Our empirical results seem very promising and show the utility of our methodology with regard to the intelligibility and actionability of an SVM output.
- CCLS-08-02.pdf application/pdf 614 KB Download File