(Day 97) Review of the GNN structure and training (last 2 days) + starting Colab 2 of XCS224W - ML with Graphs

Ivan Ivanov · April 7, 2024

Hello :) Today is Day 97!

A quick summary of today:

  • read through and reviewed my notes on intro to GNNs (Day 94)
  • read through and reviewed my notes on Designing a GNN layer (Day 95)
  • read through and reviewed my notes on GNN training (Day 96)
  • started 2nd homework of the course covering the above two

First I had a quick look through the 2nd homework that covers material covered in Day 94 and Day 95, and I decided that it would be best to do a proper review of the important concepts from those days

Topics I went over:

Designing a single layer of a GNN, message computation, aggregation, GCN, GraphSAGE, GAT, attention and multi-head attention in graphs, stacking GNN layers, the problem of over-smoothing, shallow GNNs, using skip connections graph augmentation, feature augmentation, training GNNs on a node-level, edge-level and graph-level, pooling for graph-level tasks, DiffPool, supervised, unsupervised and self-supervised learning, loss functions, evaluation metrics, splitting data

As for assignment two, I am not allowed to share any code from it, but today I did not finish it. I decided to go slowly with this one and completed just 4 of the questions. Tomorrow I think I will be able to do it all and submit to get a final grade ^^

That is all for today!

See you tomorrow :)