A Deep Orthogonal Non-negative Matrix Factorization Method for Learning Attribute Representations

摘要

Orthogonal non-negative matrix factorization (ONMF) is a powerful unsupervised learning method because it is equivalent to the K-means method and can be more robust and flexible for clustering analysis. Arguing that ONMF with a single layer implementation often fails to capture the potential hierarchical features of complex objects, a deep orthogonal NMF (deep ONMF) model with cascaded multiple ONMF layers was proposed in this paper. We demonstrated how deep ONMF is able to reveal the hierarchy information of data and hence lead to improved clustering performance by both theoretical analysis and experiments on real-world data.