ABSTRACT

A graph of interconnected nodes represents data that includes relationships. Social network analysis is the study of such graphs that looks at questions about their structures and patterns in order to comprehend the data better and forecast social network trends. Static analysis, which ignores the time of interaction (i.e., the network is frozen in time), fails to capture the evolutionary patterns in dynamic networks. Detecting community evolutions, or changing community structures over time, offers information about the network's underlying behavior. Several academics have recently begun focusing on finding important events that characterize community evolution in dynamic environments.

The graph generation is a crucial problem in understanding the evolution of a network connecting different entities, for example, fraud detection, malicious nodes analysis, and generating different communities in the network. In this work, we propose an unsupervised graph generation approach that learns the spatial topographical structure of social networks and generates a temporal graph on the basis of previously generated nodes. Modeling social interactions, identifying novel chemical structures, and generating knowledge graphs are all examples of essential applications for generative models for real-world graphs. The GNN-based approach learns node embeddings for each graph separately and uses attention techniques to match them. When compared to traditional relaxation-based approaches, this method performs well in practice and in real-world scenarios.

In this work we propose a three-step process to generate the graph in an unsupervised manner, the nodes and topological information is captured and embedded using Node2Vec based biased random walks and these node embeddings are used to generate edges using a variational autoencoder based model, and the latent representation is used to connect node embedding via learning the explicit representation of the network, now the generation of the graph is done in a sequential manner using recurrent neural networks with the help of previously obtained embedding of nodes and edges, in this manner the generated graph contains similar properties of the network but different structural information, the resultant graph exhibit temporal evolution of the existing graph. The empirical analysis of the results obtained graph depicts the 90% structural similarity between graphs, achieving state-of-the-art results on the evolution models.