Evolving-Graph Gaussian Processes
Speaker: David Blanco Mulero
Email: david.blancomulero(at)aalto.fi
Robotics Seminar Series. Fifth Session – 15th October 2021, 15:30-16:00, via zoom (use this link): https://aalto.zoom.us/j/62124942899
Abstract: Graph Gaussian Processes provide a data-efficient solution on graph structured domains. However, existing approaches assume no temporal evolution of the graph structure, whereas many real graph data present an evolving relationship.In this talk, I will present our recent work “evolving-Graph Gaussian Processes”.
Our work proposes a new Gaussian Process (GP) model that learns the transition of graph vertices. Our approach measures the similarity between vertices via a neighbourhood kernel, which models the interaction changes between vertices. We assess the performance of our method on simulated physical systems where graphs evolve over time, and demonstrate the benefits of our model over static graph GP approaches.
References:
Blanco Mulero, David, Markus Heinonen, and Ville Kyrki. “Evolving-Graph Gaussian Processes.” International Conference on Machine Learning, Time Series Workshop. 2021.