Graph-structured data is ubiquitous and occurs in several application domains. Graph representation learning approaches, however, have been limited to such applications where a graph structure is given or have used heuristics to construct affinity graphs before learning commences. One of the long-standing goals of machine learning is to infer and leverage relational dependencies, even if not available a-priori. We propose an end-to-end differentiable framework that jointly learns the graph and weights of a graph convolutional network by approximately solving a bilevel program whose inner and outer objectives aim at optimizing, respectively, the parameters of a GCN and its graph structure. This makes graph neural networks applicable to a much wider range of learning problems. We show that the proposed method outperforms related approaches by a significant margin on datasets where the dependency structure is either incomplete or completely missing. I'll also describe some applications of graph neural networks in the (bio-)medical domain.
Mathias Niepert is a chief research scientist of the Systems and Machine Learning (SysML) group at NEC Labs Heidelberg. From 2012-2015 he was a postdoctoral research associate at the Allen School of Computer Science, University of Washington. He was also a member of the Data and Web Science Research Group at the University of Mannheim and co-founded several open-source digital humanities projects such as the Indiana Philosophy Ontology Project and the Linked Humanities Project.
The event will take place on Tuesday, 9 July, 2019 at 6pm at the DKFZ Communication Center (K1+K2), Im Neuenheimer Feld 280. Drinks and snacks will be provided, courtesy of the Division of Medical Image Computing at DKFZ. Kindly help us plan ahead by registering for the event on our meetup page.