Stanford CS224W: Machine Learning with Graphs | 2021 | Lecture 8.2 - Training Graph Neural Networks
36 هزار بار بازدید -
3 سال پیش
-
For more information about Stanford’s
For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: https://stanford.io/3mgWjr5
Jure Leskovec
Computer Science, PhD
Now that we have discussed methods for augmenting graphs to improve graph representations, we will talk about methods for training GNNs. Specifically, we present a GNN training pipeline. Previously, we have discussed the first three components of the pipeline: 1) representing data as a graph, 2) GNNs as neural network models over graphs, and 3) using GNNs to generate node embeddings. In this part of the lecture we will discuss the next components of the GNN training and evaluation pipeline: 4) generating predictions for different graph tasks based on node embeddings (prediction heads), 5) using model predictions and labels to a) define a loss function and b) compute evaluation metrics. Through this process we will cover GNNs applied to node-level, edge-level, and graph-level tasks, discuss both supervised and unsupervised methods for training, and highlight DiffPool as a powerful GNN based model for graph classification.
To follow along with the course schedule and syllabus, visit:
http://web.stanford.edu/class/cs224w/
Jure Leskovec
Computer Science, PhD
Now that we have discussed methods for augmenting graphs to improve graph representations, we will talk about methods for training GNNs. Specifically, we present a GNN training pipeline. Previously, we have discussed the first three components of the pipeline: 1) representing data as a graph, 2) GNNs as neural network models over graphs, and 3) using GNNs to generate node embeddings. In this part of the lecture we will discuss the next components of the GNN training and evaluation pipeline: 4) generating predictions for different graph tasks based on node embeddings (prediction heads), 5) using model predictions and labels to a) define a loss function and b) compute evaluation metrics. Through this process we will cover GNNs applied to node-level, edge-level, and graph-level tasks, discuss both supervised and unsupervised methods for training, and highlight DiffPool as a powerful GNN based model for graph classification.
To follow along with the course schedule and syllabus, visit:
http://web.stanford.edu/class/cs224w/
3 سال پیش
در تاریخ 1400/02/15 منتشر شده
است.
36,089
بـار بازدید شده