A comparison of network embedding approaches

Network embedding automatically learns to encode a graph into multi-dimensional vectors. The embedded representation appears to outperform hand-crafted features in many downstream machine learning tasks. There are a plethora of network embedding approaches in the last decade, based on the advances and successes of deep learning. However, there is no absolute winner as the network structure varies from application to application and the notion of connections in a graph has its own semantics in different domains. In this report, we compare different network embedding approaches in real and synthetic datasets, covering different graph structures. Although our prototype currently includes only two network embedding techniques, it can be easily extended due to our systematic evaluation methodology, and available source code.