Deep Learning for Network Analysis: Problems, Approaches and Challenges
The analysis of social, communication and information networks for identifying patterns, evolutionary characteristics and anomalies is a key problem for the military, for instance in the Intelligence community. Current techniques do not have the ability to discern unusual features or patterns that are not a priori known. We investigate the use of deep learning for network analysis. Over the last few years, deep learning has had unprecedented success in areas such as image classification, speech recognition, etc. However, research on the use of deep learning to network or graph analysis is limited. We present three preliminary techniques that we have developed as part of the ARL Network Science CTA program: (a) unsupervised classification using a very highly trained image recognizer, namely Caffe; (b) supervised classification using a variant of convolutional neural networks on node features such as degree and assortativity; and (c) a framework called node2vec for learning representations of nodes in a network using a mapping to natural language processing.