Unsupervised Dimension Reduction Using Supervised Orthogonal Discriminant Projection for Clustering
Leilei Yan, Li Zhang · 2019
This paper proposes a novel unsupervised dimensionality reduction method for clustering by combining supervised orthogonal discriminant projection (SODP) and K-means selective clustering ensemble, called SODP-KSCE. The novel algorithm, operating in an iterative manner, adaptively optimizes the clustering results and learns a subspace with optimal separation. To enhance the stability of K-means, SODP-KSCE adopts ensemble learning. Moreover, a negentropy increment (NI) index is introduced to measure the clustering performance. The K means clustering ensemble algorithm is performed in the low-dimensional subspaces to generate pseudo class labels for unlabeled data, which are then adopted to guide the dimension reduction process of SODP in the original space. Experimental results on multiple data sets indicate the effectiveness of SODPKSCE.