Constrained Multi-view NMF with Graph Embedding for Face Clustering

Abstract

Face clustering that aims to group faces from the same people is a key component in face tagging and attribute analysis. Nonnegative matrix factorization (NMF) has shown competitiveness for clustering, but lacks of discrimination in practical tasks. In this paper, we propose a constrained multi-view NMF method with graph embedding (GCMNMF) for face clustering. GCMNMF incorporates the graph constraint and label constraint into an unified framework. GCMNMF aims to seek latent discriminative representations for multiple views, and maintain the within-view geometric structure simultaneously. In addition, an iterative optimization algorithm based on multiplicative rules is developed to efficiently solve GCMNMF. Experimental results on two real-world datasets demonstrate the effectiveness of the proposed method on face clustering tasks.