Support Vector Machine Classification Trees

Proteomic and metabolomic studies based on chemical profiling require powerful classifiers to model accurately complex collections of data. Support vector machines (SVMs) are advantageous in that they provide a maximum margin of separation for the classification hyperplane. A new method for constructing classification trees, for which the branches comprise SVMs, has been devised. The novel feature is that the distribution of the data objects is used to determine the SVM encoding. The variance and covariance of the data objects are used for determining the bipolar encoding required for the SVM. The SVM that yields the lowest entropy of classification becomes the branch of the tree. The SVM-tree classifier has the added advantage that nonlinearly separable data may be accurately classified without optimization of the cost parameter C or searching for a correct higher dimensional kernel transform. It compares favorably to a regularized linear discriminant analysis, SVMs in a one against all multiple classifier, and a fuzzy rule-building expert system, a tree classifier with a fuzzy margin of separation. SVMs offer a speed advantage, especially for data sets that have more measurements than objects.