Institute of Information Science, Academia Sinica

Although the generalization power of (axis-parallel) decision tree can be compromised by the strict requirement of rectangular partition of a data space, we show that decision tree can be used to accelerate the training and testing of support vector machines (SVMs). In the training process, we use a decision tree to decompose a given data space and train SVMs on the decomposed regions. In the testing process, we are able to use a number of decision trees to derive a reduced set of class types for a given test sample. We apply the above two accelerating techniques to data sets of various sources and demonstrate that we can substantially trim down the times in training and testing linear or non-linear SVMs, and still achieve comparable test accuracy.