Abstract
Subspace clustering is a critical unsupervised learning task for high-dimensional data, assuming that data points lie in a union of low-dimensional subspaces. Traditional methods like Low-Rank Representation (LRR) often suffer from rigid regularization and sensitivity to noise. This paper introduces Correlation Self-Expression Shrunk (CSS), which utilizes a compound Schatten p-norm and a robust shrunk pattern with adaptive loss. This approach effectively captures both sparse and dense structures while maintaining resilience to outliers. Experimental results on image benchmarks demonstrate significant performance improvements over state-of-the-art methods in both accuracy and visual clarity.
Introduction
Subspace clustering aims to group high-dimensional data points, such as images of faces or objects, into their respective low-dimensional subspaces. While Low-Rank Representation (LRR) is a popular framework for this task, it has two major weaknesses. First, LRR typically uses the nuclear norm as a convex surrogate for the rank function, which is often too rigid to induce the ideal block-diagonal structure required for clean clustering. Second, standard models rely on Euclidean distance, making them highly sensitive to outliers and noisy data points, which degrades clustering performance.
Methodology
The CSS model introduces three key technical innovations to enhance the LRR framework:
1. Compound Schatten p-norm
Instead of a single nuclear norm, CSS uses a combination of two Schatten p-norms. As illustrated in Figure 1, different values of $p$ change how the norm approximates the rank function. By setting $p < 1$ (e.g., 0.5), the model induces sparsity and a clear block-diagonal structure. Conversely, setting $p > 1$ (e.g., 2) allows it to capture denser, global correlations. Combining these allows the model to balance these behaviors adaptively.
2. Robust Shrunk Pattern
The authors incorporate a "pattern shrinking" strategy. Rather than just learning a self-expression matrix $Z$, they learn a refined, "shrunk" version $U$. This $U$ is designed to be more compact, further reducing the deviation between similar samples and suppressing noise in non-block parts of the matrix.
3. Adaptive Loss
To handle outliers, the model replaces standard Euclidean distance with an adaptive loss function—a compromise between $L_1$ and $L_2$ norms. This makes the gap between the raw representation and the shrunk pattern more robust to "dirty" data.
Experimental Results
The model was evaluated on four widely-used image clustering benchmarks: JAFFE, PALM, Extended YaleB, and MSRA. CSS consistently outperformed baseline methods like LRR, Sparse Subspace Clustering (SSC), and Correlation Adaptive Subspace Segmentation (CASS) across multiple metrics.
- Clustering Accuracy: On the Extended YaleB dataset, CSS achieved 69.39% accuracy, a massive jump over LRR's 53.65%. On JAFFE, it reached a near-perfect 98.12%.
- Normalized Mutual Information (NMI): CSS showed a 2.6% improvement over the next best methods on the JAFFE and MSRA datasets, indicating superior ground-truth structure recovery.
- Visual Clarity: Qualitative results shown in Figure 2 demonstrate that CSS produces much cleaner, more distinct block-diagonal affinity matrices compared to other advanced variants, which often produce noisy entries outside the main blocks.
Conclusion
Correlation Self-Expression Shrunk (CSS) represents a significant advancement in subspace clustering by integrating flexible regularization and robust pattern shrinking. While the model introduces additional hyperparameters and computational complexity, its ability to produce clean affinity matrices and handle noisy data makes it a powerful tool for high-dimensional unsupervised learning. Future research may explore automated parameter tuning or extending these concepts into deep neural network architectures.
Summary updated