Abstract: Aiming at the problem that the deep embedding clustering method only considerd the intra-class relationship and multi-view clust
ering had insufficient feature representation, we proposed a deep embedding clustering method based on inter-class loss and multi-view fusion. The method introduced a new regularization term into the loss function of deep embedding clustering to improve the class discrimination. Firstly, the feature representations of multi-view data were extracted by the auto-encoder, and the feature representations of different views were fused to obtain the public representation. Based on this, the soft allocation distribution and auxiliary target distribution of data were obtained. Secondly, the final clustering result was obtained by jointly optimizing the public representation and cluster allocation. The experimental results on multi-view datasets show that this method can effectively improve the clustering performance.

Key words: deep learning, multi-view clustering, feature fusion, inter-class loss