Improving clustering performance using independent component analysis and unsupervised feature learning

Abstract

ObjectiveTo provide a parsimonious clustering pipeline that provides comparable performance to deep learning-based clustering methods, but without using deep learning algorithms, such as autoencoders.Materials and methodsClustering was performed on six benchmark datasets, consisting of five image datasets used in object, face, digit recognition tasks (COIL20, COIL100, CMU-PIE, USPS, and MNIST) and one text document dataset (REUTERS-10K) used in topic recognition. K-means, spectral clustering, Graph Regularized Non-negative Matrix Factorization, and K-means with principal components analysis algorithms were used for clustering. For each clustering algorithm, blind source separation (BSS) using Independent Component Analysis (ICA) was applied. Unsupervised feature learning (UFL) using reconstruction cost ICA (RICA) and sparse filtering (SFT) was also performed for feature extraction prior to the cluster algorithms. Clustering performance was assessed using the normalized mutual information and unsupervised clustering accuracy metrics.ResultsPerforming, ICA BSS after the initial matrix factorization step provided the maximum clustering performance in four out of six datasets (COIL100, CMU-PIE, MNIST, and REUTERS-10K). Applying UFL as an initial processing component helped to provide the maximum performance in three out of six datasets (USPS, COIL20, and COIL100). Compared to state-of-the-art non-deep learning clustering methods, ICA BSS and/or UFL with graph-based clustering algorithms outperformed all other methods. With respect to deep learning-based clustering algorithms, the new methodology presented here obtained the following rankings: COIL20, 2nd out of 5; COIL100, 2nd out of 5; CMU-PIE, 2nd out of 5; USPS, 3rd out of 9; MNIST, 8th out of 15; and REUTERS-10K, 4th out of 5.DiscussionBy using only ICA BSS and UFL using RICA and SFT, clustering accuracy that is better or on par with many deep learning-based clustering algorithms was achieved. For instance, by applying ICA BSS to spectral clustering on the MNIST dataset, we obtained an accuracy of 0.882. This is better than the well-known Deep Embedded Clustering algorithm that had obtained an accuracy of 0.818 using stacked denoising autoencoders in its model.ConclusionUsing the new clustering pipeline presented here, effective clustering performance can be obtained without employing deep clustering algorithms and their accompanying hyper-parameter tuning procedure.