Univariate Filter Technique for Unsupervised Feature Selection Using a New Laplacian Score Based Local Nearest Neighbors

Knowing the actual relevant features of a given data set can speed up the learning and classification processes. Most of the studies on feature selection techniques concern the classification in supervised learning. Very few studies focus on unsupervised classification. However, selecting the relevant features in unsupervised learning is more difficult than supervised learning since the topology of the given data space must be strongly preserved. These are a few proposed techniques, especially filter technique based on Laplacian score, for unsupervised feature selection. However, these techniques concern only local topology of the data clusters. In this paper, a new univariate filtering technique, called Laplacian++, is proposed and based on the strong constraint on the global topology of the data space. We apply Laplacian++ to several public datasets of UCI repository of machine learning databases and compare its performance with Laplacian score. The experimental results signify that the performance of our proposed technique is obviously better than those from the other techniques.