Joint Deep Recurrent Network Embedding and Edge Flow Estimation.

The two most important tasks of network analysis are network embedding and edge flow estimation. The network embedding task seeks to represent each node as a continuous vector, and the edge flow estimation seeks to predict the flow direction and amount along each edge given some known flows of edges. In the past works, they are always studied separately, while their inner connection are completely ignored. In this paper, we fill this gap by building a joint learning framework for both node embedding and flow amount learning. We firstly use a long short-term memory network (LSTM) model to estimate the embedding of a node from its neighboring nodes’ embeddings, meanwhile we use the same LSTM model with a multi-layer perceptron (MLP) to estimate a value of the node which presents its importance over the network. The node value is further used to regularize the edge flow learning, so that for each node the balance of flowing-in and flowing-out reach the node value. We simultaneously minimize the reconstruction error of neighborhood LSTM for each node, the approximation error of node value, and the consistency loss between node value and its conjunctive edge flow values. Experiments show the advantage of the proposed algorithm over benchmark datasets.