PROJECT TITLE :

Role-based Graph Embeddings

ABSTRACT:

The concept of random walks is at the core of a good number of the currently available methods for node embedding and network representation learning. However, these methods suffer from a number of drawbacks that are caused by the utilization of conventional random walks. For instance, the embeddings that are produced by these methods capture the proximity (communities) among the vertices, rather than the structural similarity between the vertices (roles). Because embeddings are associated with the identities of individual nodes, it is not possible to move them to different nodes or graphs. We introduce the Role2Vec framework to learn structural role-based embeddings. This framework is based on the proposed notion of attributed random walks, which we hope will allow us to overcome these limitations. Particularly noteworthy is the fact that the framework can be used as a foundation for generalizing any walk-based method. By learning inductive functions that capture the structural roles in the graph, the Role2Vec framework makes it possible for these methods to be applied to a wider range of contexts. When each vertex is mapped to its own function that unmistakably identifies it, the original methods are recovered as a special case of the framework. This happens when the original methods were first implemented. Finally, it has been demonstrated that the Role2Vec framework is effective, with an average AUC improvement of 17.8 percent for link prediction, while requiring on average 853 times less space than existing methods on a variety of graphs from a variety of domains. This was achieved by testing the framework on different graphs from each of these domains.