Graphs, Vectors and Machine Learning - Computerphile

There's a lot of talk of image and text AI with large language models and image generators generating media (in both senses of the word) - but what about graphs? Dr David Kohan Marzagao specialises in Machine Learning for Graph-Structured Data and takes us through some simple examples.


This video was filmed and edited by Sean Riley.

[architech]

If all computer science lectures were presented like this gentleman did in this video, all students would excel without a doubt. Great job on this video.

[bottomhat]

In case anyone is a bit lost as to what's going on with the dot product. Basically, it's a way of comparing two vectors for similarity. So if you've got two identical vectors of length 1 - so both pointing the same way - the dot is 1. Meaning identical.

Turn one round 180 degrees and the dot product gives -1. If it's perpendicular you get zero. It's a lot like the cosine if you remember your trig.

Bit more to it if they're not unit vectors with a length of one. It basically gives you the ratio if one were projected onto the other. Imagine a clock face with the long hand at 12. The dot product gives you the amount the short had goes up the long hand as a ratio. So if you shine a light from the side and the shadow of one hand goes half way up the other; then the dot product is 0.5.

Hope that helps.

[viktortodosijevic]

I like this. I'd watch more of this guy explaining graph kernels and beyond!

[j]

Finally something I can understand!!
I hate other explanations either asking to ignore the math, or immediately dive into the code. As a beginner in computer science, I have no idea how different concepts get connected. This guy helps bridge that missing link in the knowledge for me!🙏🙏😆
Thank you so much!! Can he explain also the Kernel trick!

[adityavardhanjain]

This guy just oozes passion for computer science!

[danielwtf]

Really interesting. For all interested in those directions I recommend researching in the direction of graph neural networks and specifically for molecules topological neural networks!

[novelspace]

All this leads into the SVD where the U matrix captures the eigen vectors of the AAT matrix(observations) and the V matrix captures the eigenvectors of the ATA matrix(features)

[kabuda]

Nice presentation. More content on graphs optimizations

[wouldntyaliktono]

You can use graph convolutional networks to classify certain kinds of abuse by the users of your product. It's really flexible when you have sparse bits of information about the relationships between people and products. Graphs are absolutely everywhere.

[Amonimus]

I'd think the similarity between two objects is the % of overlapping arrangements. Identical objects would have 1 similarity, while objects that share a large pattern of elements would have about 0.5 similarity. The problem is that going through all those arrangements would get astronomically many right away.

[odorlessflavorless]

So, it seems Dr David is my favorite now :)

[YashGoswami-yi]

I think we need more videos on Graph theory.

[AuditorsUnited]

so text language models use a mapping coordinate system to find what words are most likely near it.. it would work with the graph triangle demos easily

[mellertid]

A general computerphile comment: work is slowly being done to build Babbage's analytical engine. Maybe you can feature the project and give it some momentum!

[JamesGaehring]

The "plots", distinguished from graphs, which Sean raises at 2:30, are those plots as in a Cartesian coordinate system, like a grid with x and y axes? I assume these algorithms don't apply to plots b/c coordinates are positional, ie, the distance, angle, orientation or other scalar properties of any given coordinate is fixed to the plane itself, and not just relative to other nodes in a graph. Is that more or less why?

[beeble]

13:52 It's Leman, not Lehman. Andrei Leman was Russian, not German. But probably half the papers in the literature make this mistake! His co-author, Boris Weisfeiler, at some point emigrated to the US and disappeared in Chile in mysterious circumstances during the Pinochet dictatorship.

[writerightmathnation]

@08:35, “We’re not going to go into the details of the kernel method…”.
Do you have a recommendation for someone who wants to go into the details? A reference maybe?

[owenhoffend]

What happens when you use this technique with a distance of 3 and you start having to consider loops?

[bmitch]

Instead of three red and three blue nodes, what if we had three blue and one brown? 😁

[kabrol]

Is it not feasible to use an encoder-decoder model to build a low dimension embedding of an adjacency matrix, and then perform cosine similarity?