Spectral clustering ensemble via compositional data clustering

An unsupervised learning algorithm for spectral clustering (SC) ensemble is proposed in this paper. A new consensus function is designed to combine multiple spectral clusterings. The proposed algorithm is suitable to large-scale data (e.g., texture image) and it can solve the sensitivity of scaling parameter of spectral clustering. The random scaling parameter and Nyström approximation are used to generate the individuals of SC for ensemble learning and the generated labels are regarded as the new features for each sample. Hungarian algorithm is used to realign the labels and then a compositional data vector can be found by computing the ratio of each label for data points. The compositional data vectors are mapped into another space via logcontrast transform to solve the ill-posed problem of compositional data and final ensemble result can be achieved by clustering the mapped data. Experimental results on UCI data and texture images show that, by comparison of previous approaches based on hypergraph and mixture model, the proposed algorithm takes the least computation time with the almost identical accuracy, and it avoids the selection of accurate scaling parameter in spectral clustering.