Talk:Graph neural network

Proposed merge of Draft:Graph neural network into Graph neural network

Draft contains more information than article. Robert McClenon (talk) 05:34, 7 July 2021 (UTC)

Oh wow, yeah your article is far more detailed. I second this Zaya (talk) 16:17, 7 July 2021 (UTC)

The tone of the draft is practitioner oriented, as seen here:

On the other hand, GNN not only be able to applied to non-Euclidean data but also Euclidean data such as sentences, images or videos since such data can be represented as graph data if organized properly.

That's effectively speculative language lifted right out of a funding request (more research required).

On the other hand, the following encyclopedic claim from the current stub article is not found in the draft:

It has been mathematically proven that GNNs are a weak form of the Weisfeiler–Lehman graph isomorphism test, so any GNN model is at most as powerful as this test.

On the basis of these surface details, I'm concerned that the draft is so thoroughly practitioner oriented as to be somewhat unbalanced. It's not the point of Wikipedia to document GNNs as a happening thing. On the other hand, I applaud the tremendous expansion of coverage the draft seems to embody (I'm not expert enough to decide this for myself).

Observations and suggestions for improvements

The following observations and suggestions for improvements were collected, following expert review of the article within the Science, Technology, Society and Wikipedia course at the Politecnico di Milano, in June 2022.

In the following some suggestions to improve the article.

- In the introduction, the sentence: "As of 2022, whether is it possible to define GNN architectures "going beyond" message passing, or if every GNN can be built on message passing over suitably defined graphs, is an open research question.[8]" is too generic and should be elaborated a bit by briefly explaining why there is the need for going beyond message passing.

- In the Architecture section, the period "It has been demonstrated that GNNs cannot be more expressive than the Weisfeiler–Lehman Graph Isomorphism Test.[20][21] In practice, this means that there exist different graph structures (e.g., molecules with the same atoms but different bonds) that cannot be distinguished by GNNs. More powerful GNNs operating on higher-dimension geometries such as simplicial complexes can be designed.[22] As of 2022, whether or not future architectures will overcome the message passing primitive is an open research question.[8]", is in my opinion out of the context of the architecture section and should be better introduced.

- In the Graph Attention section,. A better introduction to attention should be provided. The sentence "how important node u.." should be elaborated a bit. --BarettoDiArchitettura (talk) 08:44, 21 July 2022 (UTC)

The current stub is woefully inadequate, so in my mind the question is not whether to merge, but how to merge. First of all, the matter is to decide the right encyclopedic stance, and then I think the best of both worlds will easily fall into place. — MaxEnt 21:51, 24 July 2021 (UTC)

I am new to editing a wiki article and my profession is actually researching about deep learning in computer vision, so indeed my writing style was more like a computer-vision-oriented review paper, and I am not sure if this is proper for a wiki article.

The ICLR paper related to this sentence:

It has been mathematically proven that GNNs are a weak form of the Weisfeiler–Lehman graph isomorphism test, so any GNN model is at least as powerful as this test.

can be placed after GAT in the draft as another subgraph since the model proposed in the paper (GIN) is also a spatial approach. And the rest of the current article can be easily added to the introduction of the draft. Since I only write things that I have surveyed during my research, most methods I mentioned are computer-vision-oriented and hence somehow unbalanced. In my opinion, my draft is just a start, and it surely requires more experts in different areas who are using GNNs such as social networks, citation networks, biological or chemistry science fields to help edit this artical. Sheng-Yu PJ Huang (talk) 03:51, 25 July 2021 (UTC)

Content created between 28 June and 2 July 2021 has been history-merged – feel free to recover any of this content to the current version of the article. – wbm1058 (talk) 14:50, 25 October 2021 (UTC)

Overview of the article

A very detailed article, with a good math formalism to explain how to represent and implement GNNs. You may contact me at my talk page.

Review of the article

SlipherD (talk) 12:58, 1 July 2022 (UTC) -- The article is very well done and explains GNNs in deep detail, covering also some interesting applcations. My edits regarded only the use of the word "where" after a formula, which should - in general - not be capitalized. Also, but I would like confirmation on this, perhaps the $a$ just before the end of the subsection Graph attention network should be an $\mathbf {a}$? Good work and thank you!

Thank you for the style improvement! The $\mathbf {a}$ is indeed a vector and should be made bold. Thank you for pointing that out! NickDiCicco (talk) 13:09, 1 July 2022 (UTC)

PhD course review

The article is very detailed and well written. I would only give more context where you show the Normalized attention coefficients (i.e. why this formula is relevant, list and explain all terms). Always on this formula, you are using a specific activation function, I don't get whether I must use only that one. Beside this, well done --Aandre156 (talk) 14:06, 1 July 2022 (UTC)

Thank you very much for your feedback!

In the paper by Veličković et al. it is reported that coefficients are normalised such that it is easier to compare them between different nodes. I will add this statement to the subsection. Thank you!

Regarding the choice of the LeakyReLU, this was also reported in the original paper of Veličković et al. It is also my opinion that, in general, one could use any activation function. Since in the subsection I am reporting the original formulation by Veličković et al. I reckoned I should stick with that. NickDiCicco (talk) 14:17, 1 July 2022 (UTC)