Stanford CS224W: ML with Graphs | 2021 | Lecture 16.1 - Limitations of Graph Neural Networks

For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: https://stanford.io/3nDnFHr

Jure Leskovec
Computer Science, PhD

In this lecture, we will talk about advanced GNN topics. We will first discuss the limitations of the Graph Neural Networks that we have introduced so far. We summarize 2 main imperfections of existing GNNs. First, existing GNNs will always fail on certain position-aware tasks, where we want to embed nodes based on their positions in the graph rather than their neighborhood structure; the solution we will introduce is Position-aware Graph Neural Networks. Second, the message passing GNNs we have introduced have expressive power upper bounded by the WL test; we will discuss how to overcome this limitation by introducing Identity-aware Graph Neural Networks.

To follow along with the course schedule and syllabus, visit:
http://web.stanford.edu/class/cs224w/

#machinelearning #machinelearningcourse

Видео Stanford CS224W: ML with Graphs | 2021 | Lecture 16.1 - Limitations of Graph Neural Networks канала Stanford Online