Knowledge Extraction from Support Vector Machines: A Fuzzy Logic Approach

Support vector machines (SVMs) proved to be highly efficient computational tools in various classification tasks. However, SVMs are nonlinear classifiers and the knowledge learned by an SVM is encoded in a long list of parameter values, making it difficult to comprehend what the SVM is actually computing. We show that certain types of SVMs are mathematically equivalent to a specific fuzzy–rule base, called the fuzzy all–permutations rule base (FARB). The equivalent FARB provides a symbolic representation of the SVM functioning. This leads to a new approach for knowledge extraction from SVMs. An important advantage of this approach is that the number of extracted fuzzy rules depends on the number of support vectors in the SVM. Several simple examples demonstrate the effectiveness of this approach.