Deep K-Means: A Simple and Effective Method for Data Clustering

Clustering is one of the most fundamental techniques in statistic and machine learning. Due to the simplicity and efficiency, the most frequently used clustering method is the k-means algorithm. In the past decades, k-means and its various extensions have been proposed and successfully applied in data mining practical problems. However, previous clustering methods are typically designed in a single layer formulation. Thus the mapping between the low-dimensional representation obtained by these methods and the original data may contain rather complex hierarchical information. In this paper, a novel deep k-means model is proposed to learn such hidden representations with respect to different implicit lower-level characteristics. By utilizing the deep structure to conduct k-means hierarchically, the hierarchical semantics of data is learned in a layerwise way. The data points from same class are gathered closer layer by layer, which is beneficial for the subsequent learning task. Experiments on benchmark data sets are performed to illustrate the effectiveness of our method.