Graph neural networks
Learn message passing on graphs and build models for nodes, edges and whole graphs without leakage
A taste of a lesson
My link prediction model gets near perfect test scores. Should I be suspicious?
Yes, near perfect link prediction is a classic sign of leakage. The most common cause is that the edges you test on are still present in the graph the model uses for message passing, so a node can literally see the link it is asked to predict. Rebuild the split: hold out test edges, remove them from the training graph, and sample negative edges the same way for training and testing. If your data has timestamps, split by time as well. Then rerun with a simple baseline such as common neighbours. Quick check: in your current setup, are test edges included in the adjacency matrix during training?
Written by the teacher as an example. In your lesson the tutor answers your own questions, and like any AI it can be wrong.
What you will be able to do
- Represent a graph with adjacency and feature matrices and choose a task type
- Compute one round of message passing and a normalised graph convolution by hand
- Compare convolution, sampling and attention based graph layers
- Diagnose oversmoothing and over squashing and choose remedies
- Design graph data splits that avoid leakage and test against simple baselines
Lesson plan
- 1 Graphs as data Represent nodes, edges and features in matrices and identify the prediction task. Start
- 2 Message passing by hand Compute aggregation and update steps for a small graph. Start
- 3 Graph convolution, sampling and attention Compare the main layer families and their trade offs. Start
- 4 Readouts and graph level prediction Pool node representations into a single graph vector for tasks such as molecule properties. Start
- 5 Oversmoothing and over squashing Understand why deeper graph models often get worse and what helps. Start
- 6 Evaluation without leakage Set up transductive, inductive and time based splits that give honest results. Start
Try asking
About this tutor
For practitioners who know deep learning and have data that is naturally a network: molecules, transactions, citations, user and item interactions. You will represent graphs as adjacency structures and feature matrices, then build up message passing, where each node gathers information from its neighbours and updates its state. You compute a graph convolution step by hand, compare sampling based and attention based variants, and learn how to set up node, edge and graph level tasks. The later lessons focus on what goes wrong: oversmoothing in deep models, information squeezed through bottlenecks, and data splits that leak information across the graph.
Reviews
Students can review a tutor after a paid lesson. Nobody has yet.
About the teacher
Architectures explained from the inside: convolutions, recurrence, attention and beyond
9 tutors 362 lessons taught Sample
I teach neural network architectures and the reasoning behind them. My working life has been spent implementing models from papers, getting them to train, and finding out which details the paper forgot to mention. I like to explain an architecture by asking what problem it was built to solve and what it costs, so convolutional networks, recurrent networks, transformers and...
See Nikolai's profile and tutorsMore like this
Other tutors on the same or nearby topics.