Graphs and hypergraphs are prevalent in many real world applications arising from online social and financial platforms, recommendation systems and knowledge bases. How to represent such graph data to capture their similarities or differences between nodes and graphs? How to integrate graph data with other sources of data in representation learning? How to combine deep learning with symbolic reasoning? How to better design algorithms over graphs?
I will present a graph neural network framework for addressing these challenges based on the idea of embedding message passing algorithms into function spaces, and learning these algorithms from data. In large scale applications involving molecule design, recommendation system and knowledge reasoning, this graph neural network framework consistently achieves the-state-of-the-art results, in terms of accuracy, model size and scalability. Graph Neural Networks also appear to be a very good tool to advance AI to the next stage, which can combine deep learning with symbolic reasoning.