Deep Convolutional Center-Based Clustering

Abstract

Deep clustering utilizes deep neural networks to learn feature representation which is suitable for clustering. One popular category of deep clustering algorithms combines stacked autoencoder and k-means clustering by defining objectives including both clustering loss and reconstruction loss so that the feature representation and the cluster assignment could be learned simultaneously by the networks. However, we observe that the optimization direction of two terms are not consistent. To address this issue, in this paper, we propose a model called Deep Convolutional Center-Based Clustering (DCCBC), which replaces the usual reconstruction loss by a novel reconstruction loss based on cluster centers. The model can produce compact feature representation and capture the salient features of instances in different clusters. The optimization problem can be efficiently solved by mini-batch stochastic gradient descent and back propagation. Experiments on several benchmark datasets showed that the proposed model could achieve competitive results and outperformed some popular state-of-the-art clustering algorithms.