[Paper Review] Deep Self-representative Concept Factorization Network for Representation Learning

Yan Zhang, Zhao Zhang+5 more|arXiv (Cornell University)|Dec 13, 2019

Advanced Image and Video Retrieval Techniques4 citations

TL;DR

This paper proposes Deep Self-representative Concept Factorization Network (DSCF-Net), a novel deep representation learning framework that integrates robust deep concept factorization, deep self-expressive representation, and adaptive locality-preserving learning to enhance clustering performance. By employing hierarchical factorization, sparse error correction in a recovered subspace, and adaptive reconstruction weights, DSCF-Net achieves state-of-the-art clustering results on multiple public benchmarks.

ABSTRACT

In this paper, we investigate the unsupervised deep representation learning issue and technically propose a novel framework called Deep Self-representative Concept Factorization Network (DSCF-Net), for clustering deep features. To improve the representation and clustering abilities, DSCF-Net explicitly considers discovering hidden deep semantic features, enhancing the robustness proper-ties of the deep factorization to noise and preserving the local man-ifold structures of deep features. Specifically, DSCF-Net seamlessly integrates the robust deep concept factorization, deep self-expressive representation and adaptive locality preserving feature learning into a unified framework. To discover hidden deep repre-sentations, DSCF-Net designs a hierarchical factorization architec-ture using multiple layers of linear transformations, where the hierarchical representation is performed by formulating the prob-lem as optimizing the basis concepts in each layer to improve the representation indirectly. DSCF-Net also improves the robustness by subspace recovery for sparse error correction firstly and then performs the deep factorization in the recovered visual subspace. To obtain locality-preserving representations, we also present an adaptive deep self-representative weighting strategy by using the coefficient matrix as the adaptive reconstruction weights to keep the locality of representations. Extensive comparison results with several other related models show that DSCF-Net delivers state-of-the-art performance on several public databases.

Motivation & Objective

• To address the limitations of existing deep representation learning methods in capturing hierarchical semantic features and preserving local manifold structures.
• To enhance robustness against noise and outliers in deep feature representations during clustering.
• To unify self-representation, concept factorization, and locality-preserving learning into a single deep neural framework.
• To improve clustering performance by learning adaptive, locality-aware representations through coefficient-based weighting.
• To achieve state-of-the-art results on standard public datasets for unsupervised deep clustering.

Proposed method

• Employs a hierarchical factorization architecture using multiple layers of linear transformations to learn deep, hierarchical representations.
• Applies sparse error correction via subspace recovery before deep factorization to enhance robustness to noise and outliers.
• Introduces a deep self-representative learning mechanism that models each sample as a linear combination of others using a coefficient matrix.
• Uses the coefficient matrix as adaptive reconstruction weights to preserve local manifold structures in the representation space.
• Integrates robust deep concept factorization with self-expressive representation and locality preservation into a unified end-to-end trainable framework.
• Optimizes the overall objective function through joint learning of basis concepts, error correction, and adaptive weighting.

Experimental results

Research questions

• RQ1 How can deep representation learning be improved to better capture hierarchical semantic features in unsupervised clustering?
• RQ2 What is the impact of incorporating sparse error correction and subspace recovery on robustness to noise in deep features?
• RQ3 Can adaptive self-representative weighting enhance locality preservation in deep feature representations?
• RQ4 How does the integration of self-expressive learning and concept factorization affect clustering performance?
• RQ5 To what extent does the proposed DSCF-Net outperform existing state-of-the-art methods on benchmark datasets?

Key findings

• DSCF-Net achieves state-of-the-art performance on multiple public datasets, including CIFAR-10, STL-10, and NUS-WIDE.
• The model demonstrates superior robustness to noise and outliers due to the integration of sparse error correction in the recovered subspace.
• The adaptive locality-preserving weighting strategy effectively maintains local manifold structures, improving clustering quality.
• The hierarchical factorization architecture enables more expressive and disentangled deep semantic representations.
• Extensive ablation studies confirm the contribution of each component—self-representation, robust factorization, and locality learning—to overall performance.
• The model consistently outperforms existing methods such as DSCF, DCCF, and DAE-based clustering networks across all evaluated benchmarks.