UCL Discovery

Abstract

We present a method for training Support Vector Machines (SVM) classifiers with very large datasets. We present a clustering algorithm that can be used to preprocess standard training data and show how SVM can be simply extended to deal with clustered data, that is effectively training with a set of weighted examples. The algorithm computes large clusters for points which are far from the decision boundary and small clusters for points near the boundary. This implies that when SVMs are trained on the preprocessed clustered data set nearly the same decision boundary is found but the computational time decreases significantly. When the input dimensionality of the data is not large, for example of the order of ten, the clustering algorithm can significantly decrease the effective number of training examples, which is a useful feature for training SVM on large data sets'. Preliminary experimental results indicate the benefits of our approach.