A High-Order Clustering Algorithm Based on Dropout Deep Learning for Heterogeneous Data in Cyber-Physical-Social Systems

Abstract

An explosive growth of cyber-physical-social systems has been witnessed owing to the wide use of various mobile devices recently. A large volume of heterogeneous data has been collected from cyber-physical-social systems in the past few years. Each object in the heterogeneous dataset is typically multi-modal, posing a remarkable challenge on heterogeneous data clustering. In this paper, we propose a high-order k-means algorithm based on the dropout deep learning model for clustering heterogeneous objects in cyber-physical-social systems. We first build three dropout stacked auto-encoders, each with three hidden layers to learn the features for the different modalities of each object. Furthermore, we establish a feature tensor for each object by using the vector outer product to fuse the learned features. At last, we devise a tensor k-means algorithm to cluster the heterogeneous objects based on the tensor distance. We evaluate the proposed high-order k-means algorithm on two representative heterogeneous data sets and results imply that the proposed high-order k-means algorithm can achieve more accurate clustering results than other heterogeneous data clustering methods.