Fake News Detection using Graphs with Pytorch Geometric

[@simonelinnert8353]

Great Video, thank you very much!

[@sebbecht]

Nice job thanks for the video! and thanks for the disclaimer on the example.

[@harrydobbs4179]

Great video! I would be interested in a video about link prediction. :)

[@DeepFindr]

Hi! Thanks :) Will consider it in the future videos!

[@보라색사과-l1r]

so interesting! thanks for this wonderful video.

[@anyaegbuebuka4583]

Thanks so much for this video sir. i would really want to understand the interpretation of this model, what are the main features is it using to make prediction. i would be glad if you point me to a video or online source to read more on this. Thank you sir in anticipation.

[@datacuber]

Hi guys, really cool video. If I may ask a question. What is this .relu() you have on every layer? I had never seen something like that, and I looked it up and it is not mentioned anywhere. Thanks in advance, Alfonso

[@DeepFindr]

Relu stands for Rectified Linear Unit and is an activation function commonly used in neural networks. I recommend to read some blog posts about this on Google for further details :)

[@datacuber]

@@DeepFindr sorry if my question was not clear enough. What I really mean is that I've always seen used as nn.Relu() or F.relu() but never as a method self.conv.relu() as you used it. I searched on Pytorch docs and pytorch geometric and didn't find any references using it that way. How did you come up to that way of applying it? Thanks in advance.

[@DeepFindr]

Hi! If you use nn.relu or f.relu is just a personal choice as discussed here: discuss.pytorch.org/t/whats-the-difference-between-nn-relu-vs-f-relu/27599

[@antares1694]

Did you find the answer to why the train dataset is smaller than the test? I've encountered the same problem with every graph dataset I've used. Although, one thing I did notice is that the validation accuracy did not increase when I used the larger one as train in any of the graphs, and in a couple of cases, it even decreased.

[@DeepFindr]

Interesting. What do you mean by "every graph dataset I've used"? In pytorch geometric? This is the first time that I encountered it, but I've also only tried around 6-7 datasets so far. Maybe something is mixed up in the files that are downloaded

[@jackgolding9137]

I've also noticed this but I'm very new to GNNs

[@jonimatix]

Thanks for this. On different topic, do you know of good resources and examples to apply sequence prediction, maybe using directed graphs, like product recommendation / next event prediction?

[@DeepFindr]

Hi! Do you mean resources on time series forecasting? I found this to be very helpful: www.tensorflow.org/tutorials/structured_data/time_series Regarding product recommendation, there is a medium article: medium.com/recombee-blog/deep-learning-for-recommender-systems-next-basket-prediction-and-sequential-product-recommendation-796228b34dee Not sure if this is what you mean, maybe you can share some more details? :)

[@jonimatix]

@@DeepFindr thanks 👍was more thinking of ways to approach sequence recommendations using GNNs if that is possible?

[@DeepFindr]

arxiv.org/abs/2106.14226? arxiv.org/abs/2101.12457 Have you seen these papers ? The second one has also open-sourced code on Github :)

[@jonimatix]

@@DeepFindr awesome thanks for pointing these out 🔥

[@tejpalkumawat776]

Sir , please make a video on how can we prepare dataset like for training , testing and validation for the Link Prediction with the help of RGCN . Like take a example of knowledge graph of some text data after extracting (s,r,o) how should I prooced further ......

[@tonyli7014]

I am surprised that setting batchsize = 64 won't make the training convergent.

[@souvickdas5564]

Please make a video on software defect prediction using GNN

[@tejpalkumawat776]

Sir, what if we have a tweets and user information in the form of Excel sheet , then we want to use of Graph Neural Networks on top of it for our prediction

[@DeepFindr]

Hello :) I have a video how to build a graph dataset with pytorch geometric (GNN Project, video 2). Maybe this answers some of your questions :) Besides that, simply build the following things: X: The node feature vector per user based on the tweets. For that you need to convert the text into a feature vector with for example word2vec. Edge index: For this you need to find out who retweeted which user. Based on this you can build the edge information. You can Form a propagation graph like in the video, where each user is connected to the user that retweeted the root tweet. In the end you need to put your excel information into tensors of the following shape: X: [num users x embedding size] Edge index: [2, num edges] Best regards and good luck :)

[@ZaraZ-y4r]

hello , can you please provide me the power point slides of this video?

[@anees2410]

Alhamdulillah

[@juanete69]

Why are many straight lines in your videos not straight?

[@stanislavshubin3447]

+

[@773-o5t]

So great your video! unfortunately,while using your colab notebook,I have a problem. When excuting the code "train_data = UPFD(root=".", name="gossipcop", feature="spacy", split="train")" , there was a error "FileNotFoundError: [Errno 2] No such file or directory: './gossipcop/raw/new_spacy_feature.npz'" The only difference is that I used the cuda version of pytorch geometirc I found that folder with only directories and no data. Do you no why?

[@DeepFindr]

Hi! The data should be downloaded automatically. I tested the notebook and didn't have the error. Did you try again?

[@773-o5t]

@@DeepFindr At than time I tried serval times and failed. But now it is OK .It's too strange.

[@anyaegbuebuka4583]

@@773-o5t please what other features can i try out?