Integrating Subspace Correlations and Local Similarity: A Novel Approach to Spectral Clustering

Spectral clustering is effective for local similarity detection of low-dimensional data, adept at handling non-convex distributions and complex structures. Yet, it struggles with high-dimensional data due to the “curse of dimensionality.” In contrast, subspace clustering excels in high-dimensional spaces by identifying subspace structures but falls short with non-convex low-dimensional data. This paper presents a novel adaptive clustering algorithm that combines the advantages of spectral and subspace clustering, overcoming their individual limitations across different data dimensions. By dynamically adjusting its strategy according to data dimensionality-utilizing spectral clustering for low-dimensional data to capture local features and subspace clustering for high-dimensional data to exploit subspace relationships-the algorithm offers flexible and effective clustering for diverse datasets. Our extensive experiments on both synthetic and real-world datasets demonstrate significant improvements in clustering performance, showcasing the capability of the algorithm to adapt and excel across varying dimensions.