Extended locality preserving partial least squares with class information

Feature extraction techniques are methods widely used to reduce the dimensionality of a data while retaining most of the relevant information in the original data. Locality preserving partial least squares (LPPLS) is a recently developed feature extraction technique that aims to preserve the local structural information of data. LPPLS seeks to preserve local structure defined by nearest neighbors. However, the nearest neighbors may belong to different classes which might lead to the poor performance of LPPLS in discriminating the different classes in the data. In this paper, we propose an extension of LPPLS called extended locality preserving partial least squares which consider class label information. The binary (0-1) weighting technique together with label information is used to construct the similarity matrices that determine local projection of the data. Therefore, our extended LPPLS does not simply preserve local structure, but also has discriminating power to differentiate data from different classes. Experimental results on various data sets demonstrate the effectiveness of the proposed extended LPPLS. Two different evaluation metrics, normalized mutual information (NMI) and Fowlkes-Mallow index are used to measure the accuracy of methods used in the experiments.