Stanford CS224W: Machine Learning with Graphs - Stanford CS224W: ML with Graphs | 2021 | Lecture 9.2 - Designing the Most Powerful GNNs

For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: https://stanford.io/3nGksXo Jure Leskovec Computer Science, PhD In this lecture, we aim to design a maximally expressive GNN model. Our key insight is that a maximally expressive GNN should use the injective neighbor aggregation function that maps different neighboring node features into different embeddings. We first demonstrate neighbor aggregation functions used by popular GNN models (e.g., Graph Convolutional Networks and GraphSAGE) are not injective. We then design injective neighbor aggregation functions using neural networks and arrive at a Graph Isomorphism Network, the most powerful GNN model. To follow along with the course schedule and syllabus, visit: http://web.stanford.edu/class/cs224w/