One-step spectral rotation clustering with balanced constrains

Spectral clustering is a popular graph-based clustering method which is widely applied to pattern recognition and image segmentation. Traditional spectral clustering usually involves two separate processes: graph learning and graph-based clustering, making such a two-step strategy easily output sub-optimal performance and moreover, noisy data and imbalanced clusters could make existing clustering methods hard to meet practical necessity. To this end, one-step spectral clustering coupled with a balanced constraint is proposed to jointly optimize the robust low-dimensional representation, the spectral rotation and the cluster indicator in a unified learning framework. Specifically, dimensionality reduction is conducted by combining subspace learning and feature selection to obtain a robust low-dimensional representation. Besides, a spectral rotation mechanism is used to produce one-step clustering for reducing clustering biases, and a balanced constraint is utilized to regularize the clustering result to generate clusters with similar sizes. Moreover, an iterative optimization method is put forward to fast solve the proposed objective function. Experimental results on ten benchmark datasets compared to state-of-the-art spectral clustering methods showed that our proposed clustering method could output the balanced clusters and the competitive clustering performance simultaneously.