Instance Selection in the Projected High Dimensional Feature Space for SVM
Jianguo Liu, Riya Danait · 2018
Data classification is a supervised learning task where a training set with previously known information is used to construct a classifier. The classifier is then used to predict the class of unforeseen test instances. It is often beneficial to use a subset of the training set to construct the classifier, in particular when the size of the data set is large. For example, support vector machine (SVM), one of the most effective classifiers, only needs the support vectors to make the prediction. Therefore, all non-support vectors can be eliminated without affecting the classification performance. However, it is usually unknown which instances in the training set are support vectors before the training is completed. Researchers have developed different methods to delete the potential non-support vectors while retaining the likely support vectors before the training starts. This preprocessing to the training data set is often known as instance selection. Many of the instance selection methods are based on the geometry of the training samples. Measures in the original feature space are usually used. We propose to use measures in the projected high dimensional feature space for SVM since this is where the separating hyperplanes are determined. We compare the performance with some existing methods on a few benchmark data sets. The experiments show that using measures in the projected feature space may improve the classification accuracy, sometimes substantially.