EDCWRN: efficient deep clustering with the weight of representations and the help of neighbors

In existing deep clustering methods, it is assumed that all generated representations are equally important during the clustering procedure. However, if the model can’t learn proper cluster-oriented representations, all generated representations may not be suitable for clustering. In this case, some important representations need to be more effective than the other representations in forming optimal clusters. The existing deep clustering methods do not support this idea. Also, in most methods, Kullback–Leibler Divergence (KLD) loss function is used. KLD does not preserve global data structure. In this paper, an efficient joint deep clustering framework, termed EDCWRN, is introduced to learn representations and cluster labels, simultaneously. To overcome the mentioned problems, in EDCWRN, an automatic local representation weighting strategy is applied to weight the representations of each cluster properly. Moreover, the samples and their neighbors are involved in the learning representations procedure to generate better representation. Also, a new efficient formulation for cluster assignments is proposed. Using this formulation, global and local data structure is preserved simultaneously. Experiments show that the proposed model is more efficient than other state-of-the-art methods. The implementation-source code- of EDCWRN is made publicly available at https://github.com/Amin-Golzari-Oskouei/EDCWRN .