Label Propagation from Laplacian Score for Semi-Supervised Attribute Selection

Abstract

Attribute selection is primary for many machine learning applications. Attribute selection in labelled data is not a difficult issue, because of the presence of class labels. The correlation of the features with the class labels guides the feature selection process in supervised learning. It is a difficult issue in unlabelled data because of the absence of the labels. Most of the attribute selection algorithms are proposed for attribute selection in supervised and unsupervised learning. But in real time getting labels for all the data is much harder. So it is required to propose a feature selection technique for partially labelled data. It guides the selection of the most informative features for analysing the data. Since the data is partially labelled, the label information cannot be ignored. A semi supervised attribute selection technique known as Label Propagation based laplacian Score is proposed in this work. It works on the basis of one of the assumptions of Semi - supervised learning i.e cluster assumption. This assumption is utilized to evaluate the affinity between the data. The labels are propagated by using the label propagation technique. The features are ranked based on the laplacian score and the relevant attributes are selected based on this score.