Abstract
We propose a spectral clustering algorithm p for the multi-view setting where we have act cess to multiple views of the data, each of c which can be independently used for clusterp ing. Our spectral clustering algorithm has a w flavor of co-training, which is already a widely m used idea in semi-supervised learning. We I work on the assumption that the true underd lying clustering would assign a point to the p same cluster irrespective of the view. Hence, t we constrain our approach to only search for o the clusterings that agree across the views. d Our algorithm does not have any hyperpaS rameters to set, which is a major advantage s in unsupervised learning. We empirically b compare with a number of baseline methods t on synthetic and real-world datasets to show i the efficacy of the proposed algorithm. g t