Entropy-Constrained and Sparse Relation-Constrained Subspace Clustering

This paper proposes a novel Entropy-Constrained and Sparse Relation-Constrained Subspace Clustering (ECSSC). By integrating information entropy weighting with sparse representation, the proposed method enhances both the accuracy and robustness of high-dimensional data clustering. Key improvements include: the introduction of an information entropy weight matrix to quantify feature discriminability and improve adaptability to noise and redundant features; the use of Frobenius norm constraints on the coefficient matrix to balance computational efficiency and model performance; and the incorporation of block-diagonal constraints to refine the sparsity structure of subspaces. Experiments conducted on three image datasets—MNIST, ORL, and COIL20—demonstrate that EWSSC outperforms traditional methods, classical sparse subspace clustering algorithms, and related state-of-the-art variants in terms of clustering accuracy (ACC), normalized mutual information (NMI), and adjusted Rand index (ARI). Notably, on the ORL dataset, ECSSC achieved an NMI of 0.9031, representing a 6.18% improvement over Sparse Representation-based Clustering (SRR). Ablation studies further confirm the effectiveness of the entropy-weighting module, which contributes to an average performance gain of 10%–18% across different datasets.