Back to index
OtherAbout 60 lecture videos across the term, most of them 20 to 30 minutes·Free on YouTube with slides and Colab notebooks on the course site

CS224W: Machine Learning with Graphs

4.5

The definitive course on graph machine learning, taught by the person who wrote much of the field's foundational work. It is free, it is rigorous, and almost nothing else covers this material at this depth.

What We Liked

  • Leskovec is a primary researcher in graph representation learning, so you are learning the material from the source
  • Builds properly from classical network analysis through node embeddings to modern graph neural networks
  • Colab notebooks give you real implementation practice with PyTorch Geometric rather than pure theory
  • Lectures are cut into short, well scoped segments, which makes it far easier to work through than a course of monolithic 80 minute recordings
  • Free, with slides and reading lists published openly on the course website

What Could Be Better

  • Assumes comfort with linear algebra, probability and a working knowledge of deep learning, so it is a poor first machine learning course
  • Graph ML is a narrow speciality and this course will not help you get a general machine learning job
  • The public recordings are from earlier years, so the newest work in the field is not covered
  • Leskovec's delivery is precise but flat, and the pace can feel relentless over sixty videos
  • Homework and exam solutions stay with enrolled students, so self checking is limited

Detailed review

Graph machine learning occupies a strange position. It is genuinely important in a handful of industries, largely invisible in the rest, and almost completely absent from the standard online curriculum. Everyone learns convolutional networks and transformers. Very few people learn how to reason about data that is fundamentally relational.

CS224W is where you fix that, and it is the reason this course keeps its place in my recommendations despite being useless to most people. Jure Leskovec is not just teaching the material, he co-authored a good deal of it. GraphSAGE came out of his lab. So did a lot of the work on network motifs and large scale graph analysis that predates the current neural wave.

The consequence is that the course does not present graph neural networks as a technique dropped in from nowhere. It builds towards them. You start with classical network analysis, degree distributions, motifs, community structure, PageRank, the sort of material that was network science before anyone attached a neural network to it. Then you move to node embeddings, node2vec and its relatives, and only then does the message passing framework appear.

By the time you reach graph convolutional networks you understand what problem they are actually solving, which is a very different experience from encountering the equations cold. The structure of the recordings deserves specific credit. Rather than posting hour and a half lecture hall videos, the course is cut into segments of twenty to thirty minutes, each covering one idea. This sounds like a trivial production decision and it is not.

It changes the course from something you have to schedule around into something you can genuinely work through in the gaps of a normal week, and it makes returning to a specific concept months later actually practical. More courses should do this. The Colab notebooks are the other half of the value. They use PyTorch Geometric, which is the library you would reach for in practice anyway, and they walk you through building the models rather than just calling them.

There is a meaningful gap between having watched a lecture on graph attention networks and having written one, and the assignments close it. Now the honest caveats. This is not a beginner course and it does not pretend to be. Leskovec assumes you have the linear algebra, that you know what backpropagation is, and that a loss function does not need explaining.

Coming in without a solid deep learning foundation will not go well. More importantly, this is a speciality. If your goal is employability in general, graph ML is not where the roles are. Where it does matter it matters enormously, and the list is more interesting than people expect.

Fraud detection at payment companies is a graph problem. So is recommendation at scale. Drug discovery leans on molecular graphs heavily. Knowledge graph work has had a second life alongside retrieval augmented generation, and the intersection of graphs and language models is one of the more genuinely open areas right now.

If any of that is your world, this course pays for itself many times over. The public material is also a few years behind the frontier, which is unavoidable with recorded courses. The foundations do not go stale, but you will need to read papers to catch up on what has happened since. Leskovec's delivery is another thing to know about going in.

He is clear and extremely well organised, and he is not warm. There is no showmanship and very little levity, and over sixty videos the consistency of tone can wear. That is a matter of taste rather than a defect. My 4.5 reflects a course that does its specific job better than anything else available, at no cost, held back only by the narrowness of that job and the prerequisites required to reach it.

If graphs are relevant to you, treat this as required. If they are not, bookmark it and come back the day a problem arrives with edges in it.

[ final ]

The verdict.

If you have a graph shaped problem, whether that is fraud rings, recommendation, molecular property prediction or knowledge graphs, this is the course to take and there is no close second. If you do not, skip it until you do.