Unsupervised Extreme Learning Machine via Structured Graph Construction for Data Clustering

Abstract

Extreme learning machine (ELM) is an effective and efficient learning paradigm for multi-class classification and regression. However, most of existing ELM variants work in supervised scenario. By exploring unlabeled data based on manifold regularization, the unsupervised ELM (US-ELM) model obtained promising results on the data clustering. However, there also exist some deficiencies in US-ELM. On one hand, US-ELM is a typical two-stage process in which a data graph is formed from the data, and then optimization procedure is invoked on the fixed input data graph; on the other hand, US-ELM cannot handle the out-of-sample problem, that is, it cannot handle the new data points that are not included in the training set. To address both issues, in this paper, we propose an improved unsupervised ELM with structured graph construction (UELMSG) framework. Instead of utilizing the fixed affinity matrix to preserve the manifold structure, we construct a structured graph by adapting the affinity matrix to ELM objective in which the regression target is the clustering indicator matrix. The model formulation, optimization, complexity and convergence analysis of UELMSG are given in detail. Experiments on clustering several benchmark data sets show the effectiveness of the UELMSG.