Tensorized Non-Negative Multi-Attribute Subspace Clustering

Subspace clustering has been successful in unsupervised learning by revealing low-dimensional structures in high-dimensional data. However, most existing subspace clustering methods rely on a single self-expressive learning process and neglect the multi-attribute information inherent in real-world data. Moreover, traditional subspace clustering often fails to capture the spatial structure of the data, leading to a loss of important attribute information. To address these issues, we propose a novel approach called Tensorized Non-negative Multi-Attribute Subspace Clustering (TNMSC), which preserves both the spatial structure and the multi-attribute information of the original data. Specifically, TNMSC uses a multi-attribute learning framework to construct separate affinity representations for each attribute, while imposing non-negativity constraints to capture more interpretable and richer features. To further retain spatial information, we introduce a tensor decomposition-based representation that directly operates on the high-order (tensor) form of the data instead of flattened vectors, making our approach particularly advantageous for multidimensional image data. Finally, TNMSC explores consistent topological correlation across different attributes to learn a refined consensus affinity representation that integrates all attribute information. Extensive experiments on seven real-world datasets demonstrate that TNMSC outperforms state-of-the-art methods. For instance, on the COIL20 dataset, our method achieves a clustering accuracy of 83.71%, surpassing the second-best approach by a significant margin of 5.86%.