In this video a group of the most recent node embedding algorithms like Word2vec, Deepwalk, NBNE, Random Walk and GraphSAGE are explained by Jure Leskovec. Amazing class!
Great Sir, Congratulations for your oustanding teaching capabilities. It really change my life and my view on Graph Network. Thank you very much, Professor
Hello. These lectures are very interesting. Would it be possible to share the GitHub repositories so that I can get a better understanding of the code involved in the implementation of these concepts?
Thank you so much for making this lecture publicly available. I have a question, is it possible to apply node embedding to dynamic graphs (temporal)? Are there any specific methods/algorithms to follow? Thanks in advance for your answer.
43:40 I have a question for the slide here. How can you generalize for a new node when the model learns by aggregating the neighborhoods and the new nodes doesn't have a neighborhood yet.
Deeper networks will not always be more powerful as you may lose vector features in translation .And due to additional weight matrices the neural networks will be desensitized to feature input.Number of hidden layers should not be greater than input dimension.
In GCN, we get a single output. In GraphSAGE you concatenate it to keep the info separate. So at each step, the output H^k will have 2 outputs, isn't it? If not, then how are they aggregated and still kept separate
I think the concatenated output is the embedding of the target node. And it depends on the downstream task to further process it, by passing it through more layers, before having the final output.
ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-7JELX6DiUxQ.html "what we would like to do is here input the graph and over here good predictions will come" Yes, that is exactly it! xD