Graph SAGE - Inductive Representation Learning on Large Graphs | GNN Paper Explained
ฝัง
- เผยแพร่เมื่อ 31 ก.ค. 2024
- ❤️ Become The AI Epiphany Patreon ❤️ ► / theaiepiphany
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
In this video, I do a deep dive into the Graph SAGE paper!
The first paper that started pushing the usage of GNNs for super large graphs.
You'll learn about:
✔️All the nitty-gritty details behind Graph SAGE
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
✅ Graph SAGE paper: arxiv.org/abs/1706.02216
✅ Chris Olah on LSTMs: colah.github.io/posts/2015-08...
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
⌚️ Timetable:
00:00 Intro
00:38 Problems with previous methods
04:30 High-level overview of the method
06:10 Some notes on the related work
07:13 Pseudo-code explanation
12:03 How do we train Graph SAGE?
15:40 Note on the neighborhood function
17:40 Aggregator functions
23:30 Results
28:00 Expressiveness of Graph SAGE
30:10 Mini-batch version
35:30 Problems with graph embedding methods (drift)
40:30 Comparison with GCN and GAT
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
💰 BECOME A PATREON OF THE AI EPIPHANY ❤️
If these videos, GitHub projects, and blogs help you,
consider helping me out by supporting me on Patreon!
The AI Epiphany ► / theaiepiphany
One-time donation:
www.paypal.com/paypalme/theai...
Much love! ❤️
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
💡 The AI Epiphany is a channel dedicated to simplifying the field of AI using creative visualizations and in general, a stronger focus on geometrical and visual intuition, rather than the algebraic and numerical "intuition".
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
👋 CONNECT WITH ME ON SOCIAL
LinkedIn ► / aleksagordic
Twitter ► / gordic_aleksa
Instagram ► / aiepiphany
Facebook ► / aiepiphany
👨👩👧👦 JOIN OUR DISCORD COMMUNITY:
Discord ► / discord
📢 SUBSCRIBE TO MY MONTHLY AI NEWSLETTER:
Substack ► aiepiphany.substack.com/
💻 FOLLOW ME ON GITHUB FOR COOL PROJECTS:
GitHub ► github.com/gordicaleksa
📚 FOLLOW ME ON MEDIUM:
Medium ► / gordicaleksa
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
#graphsage #gnns #graphtheory
Any feedback is welcome - it'll help me make higher quality videos for you folks over the long run. Happy New Year! ❤
Hi, thank you. This video is realy helpful. Can you make a video about the gmmConv (gaussian mixture modele). I discovered your channel yesterday :). Have a good day.
From paris : merci.
Lovely simple illustration and comparsion with other common GNN like GCN and GAT. Really helps me. Thank you.
Really great explanation of a tough paper to understand by yourself
Thanks for this series, it has been very instructive. I will be watching the rest of the GNN playlist!
Very well explained. Even though GNNs are a new concept for me, the way you explained the background of GCN, GAT, SAGE and how Attention, LSTM are being used as aggregators, it was very easy to understand. Thank you!
That was a helpful review I really appreciate it,
Hey great explanation 🎉! Looking forward.
Please do a video on GraphSAINT
This was great!
Thank you. Great explanation :)
Brilliant. Subscribed.
Amazing video I love your content! Would love to connect and maybe talk a bit about CV + graphs for aesthetics recommendations.
I was wondering is it possible to make video with how the code implement and get the experiment results in paper, would be helpful to learn those
Loving the series. I've been interested in program synthesis ever since the ARC challenge, and I wonder why there isn't more use of GNNs on modelling programs which are inherently graphical objects, though quite sparse. Natural language description --> code, or multiline code completion would be very cool applications.
Hey Daven! Thanks for the feedback as always!
I think that Gated GNNs (Li et al.) did use GNNs for program verification and reasoning.
I think that transformers have some awesome applications when it comes to code autocompletion! The only problem is - they are just heavy! Although much of the recent work (Linformers, Longformers, etc.) focus on making them more efficient.
@@TheAIEpiphany Thanks for that reference, it rings a bell. One would think that due to the expense of programmers, this would be one of the chief interests of big tech companies, instead of occupying the fringe of academic interest.
As has been noted with GPT3 and the lack of language understanding, the massive data + transformers process will surely be able to generate plausible and syntactically correct code, but unlikely to get the semantics right so as to have the program do what you want. Or I might be surprised!
Thanks!!!
Thanks for the great explanation. I see that they are referring to two things whenever they say about the learning parameters (parameters of K aggregated functions and the set of weight matrices Wk's). I pretty much understand the learning of Wk's but didn't understand what they meant by the parameters of K aggregated functions and how are they learnt?
I got my answer. So, If we use either GCN or MEAN, the aggregator function is non-trainable but If we use POOL or LSTM, the aggregator function is trainable.
Boosting, good job!
Request to prepare video on GraphSAINT: GRAPH SAMPLING BASED INDUCTIVE LEARNING METHOD"
Hi, very helpful content. Can you please talk about how the negative samples are sampled? It's not explained in the paper and i find it very confusing.
Great explanation, Thank you so much , could you please tell me,I am really confused, in GraphSage, Do we take sample of nodes from entire graph to implement the algorithm1 or all nodes of the graph? and what difference between GraphSage and GCN as you know in GCN also embedding of any node is aggregate of embeddings on its neighborhood . I think in the graphSage algorithm also the embedding of any node is aggregate of its neighborhood . Could you tell me what is the main difference?
Awsome
Hey, great explanation! I had a question. So the way GraphSage trains, there is no way for embeddings to learn(or even be exposed to) graph structure/nodes beyond K hops right? Isn't this a shortcoming, because you might have a huge graph and maybe important structural information to learn but the formulation only allows it to see up till k hops only.
for implementing lines 1-7. After k=2, B1 has nodes from the B2 and those nodes's random neighbors from B1 right? Now when k=1, we have to find random neighbors of nodes in B1 and do union. So for this step do we find neighbors of nodes in B2 also again or only of the nodes which were originally in B1? Please help me to understand
Can anyone help me by telling me how I can use this in finding shortest path after finding node or graph classification??
Hi Aleksa, can you make a video about your background and your journey into AI/ML/DL?
Hey! Sure, I already have some snippets of information dispersed across the web (some of my YT videos, Medium blogs and an interview I recently gave to India Analytics Magazine) but I could make a dedicated video as well.
I'll definitely put that on my backlog thanks for the feedback I really appreciate it. If I see more of these I'll try and make it sooner rather than later.
👍🏽 👍🏽 👍🏽
Hey Aleksa,
could you cover the MoNet paper arxiv.org/abs/1611.08402?
The math is a bit hairy so going over it would be great! They say that their model generalizes CNN architectures to graphs and manifolds which is really important for feature learning in these domains.
Another paper on the topic of applying CNNs to graphs directly for the purpose of feature extraction is this one arxiv.org/abs/1509.09292 and it has had a huge impact in the drug modeling community, that is learning drug features/embeddings by modeling each drug as a graph and subsequently applying convolutions to these graphs. If you could go over either of these papers it would be great! I'm basically trying to understand how these "CNNs applied to graphs" differ from the GCN architecture introduced by Kipf.