The Graphics Replicability Stamp Initiative has awarded the paper «Visualizing Graph Neural Networks with CorGIE: Corresponding a Graph to Its Embedding». Congratulations to Jürgen Bernard!
The work «Visualizing Graph Neural Networks with CorGIE: Corresponding a Graph to Its Embedding» by Zipeng Liu, Yang Wand, Jürgen Bernard, and Tamara Munzner was a collaboration with the University of Zurich, the University of British Columbia (Vancouver) and Facebook.
The Replicability Stamp Award is a recognition for authors who, in addition to publishing the paper, provide a complete open-source implementation.
The proposed visual analytics solution supports machine learning experts and developers of graph neural networks (GNN) in assessing the quality of trained deep learning models and in better explaining the neural network characteristics. The principal idea is to explore correspondences between an input graph and the resulting latent space created by the GNN, to understand if GNN has learned important characteristics from the graph and to find bugs in the latent space.
This work constitutes one step towards understanding the internals of complex deep learning models and their resulting node embeddings.