Intro to Graphs and Label Propagation Algorithm in Machine Learning

[@DavenH]

Oh I'm glad I found this channel. GNNs are of particular interest to me... I think there's so much potential for neural code generation.

[@NoNTr1v1aL]

Absolutely amazing playlist! Subscribed.

[@chongtang6908]

I'm so luck I found this channel. Thank you!!!

[@KhalilMuhammad]

Amazingly clear explanation. Thanks a lot!

[@amandhaliwal3499]

So glad I found your channel

[@chinmayrath8494]

What an amazing educator. Thank you very much !!

[@CerebroneAI]

Great visualisations , thank you for the meticulous content.

[@snowjordan6822]

Keep doing videos like this.

[@jtetrfs5367]

What was the intro music, just before the main content starts at 1:30?

[@welcomeaioverlords]

Elder - Legend: kzbin.info/www/bejne/gIG1f4uknLKNpdE

[@YangJackWu]

Thank you for doing this

[@akritiupreti6974]

This makes the picture really clear. Thanks a lot! Could you also point me to any good resources on how to readily use such a technique for a very large graphs in terms of the tech stack and packages that can be used to implement this?

[@ondrejkuchta1785]

Thanks a lot!! Very helpful.

[@khim2970]

hello, appreciate your effort to make the great series of videos, and could make me clear a bit that matrix S at 6:41 is adjacency matrix, right?

[@nid8490]

Yes it is

[@devanshpurwar]

nice explanation

[@swakshardeb7908]

If we keep Y constant throughout this iterative process and initialize with random numbers for unlabeled nodes. What kind of sense does it make? Aren't we propagating wrong labels throughout the network for unlabeled nodes? Why we are not changing Y as the network keeps updating?

[@welcomeaioverlords]

I struggled with this idea too, Swakshar. But the alternative is to update Y instantly, which then causes your labeled nodes to have their real labels overwritten by its unlabeled neighbors. This leads to quick convergence to a trivial solution. I think of Y as continually pumping energy into the system as we wait for it to spread and reach a steady state. The hope is that by using something uninformative as the value for unlabeled nodes, overtime it will be dominated by true signal. And keep in mind that you can tune the contribution of this term with alpha, so you could always evaluate these trade offs.

[@swakshardeb7908]

@@welcomeaioverlords Thanks for the clarification. But why not only change the values of the unlabeled node using the previous prediction in Y and keep the true label unchanged along the iterative process. In this way, we are not overwriting the ground truth and only changing the unlabeled node information.

[@秦默雷]

I got a question about this as well. My idea about it is why we don't replace the Y with f(t), therefore it sort of become a learning process in a regular ml process.

[@artem_isakow]

Thanks a lot!

[@haroldsu1696]

Thank you Sir

[@kimminuk6042]

Is this, what's called, "Loopy Belief Propagation"?

[@PnutJpg]

pefect

[@JamesSmith-dy5vu]

Thanks, I'm excited for the series! I've seen something similar for label propagation done with a personalised pagerank algorithm. Do you know if there are many differences between the two?

[@welcomeaioverlords]

PageRank is conceptually similar: you're taking the state of a node in a graph, and sending a signal through its connections to update the state of its neighbors. And this happens iteratively until things converge. But there are also some differences. PageRank is sort of unsupervised in that there isn't a ground truth label to send, but rather it's been modeled as "the probability a random internet surfer will arrive at the page". Implementations are almost identical if you replace LP's ground truth matrix, Y, with an uninformative constant matrix of 1/N. Again, this is because PR doesn't have ground truth labels. Then, PR's damping factor d is similar to LP's alpha parameter, in that it is the relative weight between the neighbor updates and the starting signal. I hope this helps.

[@JamesSmith-dy5vu]

@@welcomeaioverlords Yes, thank you!

[@chongtang6908]

Thanks James Smith too. Also, you are so handsome, god~

[@NadavBenedek]

Good audio quality

[@henrygengiti7861]

can you please clarify sigma one more time?