Stanford CS224W: Machine Learning with Graphs | 2021 | Lecture 8.1 - Graph Augmentation for GNNs
ฝัง
- เผยแพร่เมื่อ 30 มิ.ย. 2024
- For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: stanford.io/2XQPDGQ
Jure Leskovec
Computer Science, PhD
In this lecture, we will continue talking about the different design choices when training and evaluating GNNs. Firstly, we cover graph augmentation techniques for improving the training of GNNs. We highlight two specific kinds of augmentations: 1) graph feature augmentation and 2) graph structure augmentation. For graph feature augmentation, we discuss methods for injecting additional node feature information into the graph. For graph structure augmentation, we discuss adding edges to improve message passing in sparse networks (e.g. virtual nodes), dropping edges to improve efficiency in dense networks, and subgraph sampling for reasoning over very large graphs.
To follow along with the course schedule and syllabus, visit:
web.stanford.edu/class/cs224w/
0:00 Introduction
0:31 Recap: Deep Graph Encoders
0:57 Recap: A General GNN Framework
3:58 Why Augment Graphs
5:49 Graph Augmentation Approaches
6:41 Feature Augmentation on Graphs
17:53 Add Virtual Nodes / Edges
22:35 Node Neighborhood Sampling
23:49 Neighborhood Sampling Example
The computational graph are structurally same for the node v_1 at 14:00 but they'll be fed different embeddings. The embeddings will implicitly include information about the nodes. A GNN can still struggle to distinguish cycles of varying types and shapes?
think about the same. The structures might be similar but the embedding of the additional node (4 vs 3) will still be there.
In terms of feather augumentation, why do not we use a 1-dimension value of ID as node feather?
Kim Kardashian Node ashahhahahaha
I laughed at the same thing and was about to comment 🤣🤣🤣🤣
Why node degree feature could be beneficial for us? I assume that if you assign all the nodes with ones and apply GNN you will learn node degrees pretty fast. So node degree does not contain a specific information Computational Graph cannot learn with simple initialization.