Multi-View Spectral Clustering via ELM-AE Ensemble Features Representations Learning

Lijuan Wang,Shifei Ding

doi:10.1109/access.2020.3034624

Abstract

Spectral cluster based on multi-view data has proven effective for clustering multi-source real-world data because consensus and complementary information of multi-view data ensure the result of clustering. Feature learning is the vital step in spectral clustering, and excellent feature representations can effectively improve spectral clustering performance. In this article, we propose Multi-View Spectral Clustering via ELM-Autoencoder Ensemble Features Representations Learning (MvSC-EF-ELM). First, Extreme Learning Machine as an Autoencoder(ELM-AE) learns feature representations by adopting a singular value decomposition, which reconstructs the inputs using the embedding feature space. Second, the single view and multi-view spectral clustering algorithm are applied embedding feature representations space into the eigenspace and cluster's them, respectively. Experiments on benchmark datasets demonstrate that our approach empirically validates the power of ELM-AE for feature learning from raw data and may effectively facilitate multi-view spectral clustering and induce superior clustering results.

Highlights

In the field of data mining, clustering is one of the most widely used methods of exploratory data analysis
Our proposed MvSC-embedded features (EF)-ELM algorithm builds on the normalized spectral clustering, according to Ng et al [25], and introducing the co-regularization mechanism used in the literature [8]
EXPERIMENTS AND RESULTS To verify the effectiveness of the new approach Spectral Clustering (SC)-EFELM algorithm and utilize multi-view data sets to verify the validity of the proposed MvSC-EF-ELM algorithm, we select five single-view datasets and five multi-view datasets

Summary

INTRODUCTION

In the field of data mining, clustering is one of the most widely used methods of exploratory data analysis. The latent feature space (EF) of input samples of multi-view data is learned by the ELM-AE model. To get a consistent representation of multi-view data, co-regularization is applied in the MvSC-EF-ELM model to perform the task. Single-View Spectral Clustering via ELM-AE Embedding Features Spectral clustering can obtain the result of clustering via defined an affinity matrix and constructed a Laplacian matrix and computed its eigenvalues and eigenvectors. We can select the Laplacian’s k smallest eigenvalues corresponding to the k eigenvectors to apply k-means method on it to group the data objects to k clusters These are all based on the original datasets, and feature learning majorly focuses on selecting embedded features (EF) from the input data, which could adequately describe the input datasets. Output: a clustering assignment in k clusters of the N samples in EF

Multi-view Spectral Clustering via ELM-AE Embedding Features

EXPERIMENTS AND RESULTS

Single-view Learning Results

Conclusion