Stanford CS224W: ML with Graphs | 2021 | Lecture 15.2 - Graph RNN: Generating Realistic Graphs
ฝัง
- เผยแพร่เมื่อ 30 พ.ค. 2021
- For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: stanford.io/3GBr7ec
Jure Leskovec
Computer Science, PhD
Here we provide an in-depth discussion of GraphRNN, one of the first deep generative models for graphs. The idea of GraphRNN is to generate graphs via sequentially adding nodes and edges. This graph generation process can be modeled in 2 levels: (1) node-level, where each node is added at a time; (2) edge-level, where edges are added between a new node and the existing nodes. We model these 2 processes via Recurrent Neural Networks (RNNs). After describing how an RNN can be used for generating sequences, we describe the detailed algorithm for GraphRNN training and inference. More information can be found in the paper: “GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models” arxiv.org/abs/1802.08773
To follow along with the course schedule and syllabus, visit:
web.stanford.edu/class/cs224w/
#machinelearning #machinelearningcourse
Regarding RNNs I have several questions:
1. It is known (it could be shown) that the signal in RNN would not propagate long for long sequences. Would the Teacher Forcer Help us with this common problem. Taking into account the large number of nodes in a regular graph this issue becomes important.
2. May be I am wrong, but I believe we commonly use RNNs for some sequential data where there is some correlation b/w neigbors (like with words in a sequence). However, here the nodes are independent of each other (indeed, we can choose the order of the nodes arbitrary).
22:12 - "whatever the output of that coin is the input to the next level". Is it a mistake? As we are using teacher forcing we should use real values (edge exists/or not) as I anderstand.
That's was my understanding as well!
For teacher forcing, why are we even computing for loss when we are anyways replacing the predicted probability with the real/expected value?
Only do that during training phase, thereafter use the computed values.