PROJECT TITLE :

Exploring Temporal Information for Dynamic Network Embedding

ABSTRACT:

The task of analyzing complex networks is a challenging one that is attracting an increasing amount of attention. One way to make the analysis of complex networks easier is to represent the nodes in a network as low-dimensional dense vectors. However, in the real world, networks are constantly evolving over the course of time. For example, collaboration within citation networks and communication within email networks are both examples of this. The vast majority of the more recent methods for embedding only concentrate on static networks. They do not take into account the essential temporal information, which acts as a supplement to the structural information and has been demonstrated to improve the quality of node embedding. We explore temporal information for the purpose of further enhancing the robustness of node representations in dynamic networks using a model that we call DTINE, which we propose as part of this body of work and refer to as unsupervised deep learning. In order to maintain the topology of the network, we have carefully devised a temporal weight and sampling strategy to extract characteristics from the surrounding areas. On the recurrent neural network, an attention mechanism will be applied in order to measure the contributions of historical information and capture the evolution of the networks. The results of experiments conducted on four actual networks show that the proposed method achieves higher levels of performance than the methods currently considered to be state-of-the-art.