Adaptive spectral co-clustering for multiview data

Spectral clustering is a typical unsupervised machine learning technique and it has widely adopted in various fields. This paper proposes a novel spectral clustering technique to handle the characteristics of multiview data. In the proposed method, co-training approach is adopted in the spectral clustering. When an instance has more than three views, it is difficult to handle different dependencies among views in ordinary co-training. To overcome this, the proposed method reflects these different dependencies among views when the information is propagated in the training phrase. In the experiment, the proposed method is evaluated with the synthetic data whose instances are represented with three views. The proposed method achieves up to 8.25% better ARI (Adjusted Rand Index) than those of five algorithms.