Search

Graph embedding, which constructs vector representation of nodes in a network, has shown effectiveness in many graph analysis tasks, such as node classification, node clustering, and link prediction. However, due to the complexity of graph embedding models (GEMs) and their nontransparency of hyperparameters, evaluation and comparison of embedding results in retaining the original graph features, and consequently, the selection of suitable GEMs according to graph analysis tasks are challenging for people. In this paper, we present a visual analysis method, GEMvis, to support the evaluation and comparison of GEMs from the original graph, node metric, and embedding result spaces. The method also supports the online refining of GEM by tuning the parameters in its three components (graph sampling method, neural network structure, and loss function). A series of metrics, R_node metrics, for measuring GEMs' ability to preserve specific node metrics, such as R_degree and R_closeness, is also proposed to support quantitative evaluation and comparison of GEMs' ability to preserve original graph features. Finally, three case studies and expert feedback illustrate the effectiveness of GEMvis. [ABSTRACT FROM AUTHOR]