LucaCappelletti94
commented
8 months ago Hi @nicolasdugue! Thank you for your pull request.
First of all, if I am not mistaken, erase_base_features is not defined and is never used across the model, what is up with that?
Same thing for workers.
Since this implementation as I understand it is primarily meant for didactical purposes, being it in NetworkX and all, could you kindly extend the comments and explain more extensively things such as:
- What are the parameters for, especially
erase_base_featuresandgammaand what do they do in your model? Are there optimal values? Is there any correlation between gamma and the number of nodes or the density of the graph (especially for gamma)? - How does the algorithm scale? What are the memory requirements for a graph with V nodes and E edges? Please provide both the computational and space complexities for the asymptotic worst case.
Another significant bit is testing. Please extend the test suite and test your implementation. Things such as the erase_base_features error would not have survived testing.
nicolasdugue
codecov-commenter
As fans of the karateclub package, we propose to add the implementation of SINr, a graph embedding approach based on community detection to the package. SINr was first published in : Thibault Prouteau, Victor Connes, Nicolas Dugué, Anthony Perez, Jean-Charles Lamirel, et al.. SINr: Fast Computing of Sparse Interpretable Node Representations is not a Sin!. Advances in Intelligent Data Analysis XIX, 19th International Symposium on Intelligent Data Analysis, IDA 2021. pp.325-337, ⟨10.1007/978-3-030-74251-5_26⟩. ⟨hal-03197434⟩
It allows to produce interpretable node representations and can also compute word embeddings on textual graph data : Béranger, A., Dugué, N., Guillot, S., & Prouteau, T. (2023, November). Filtering communities in word co-occurrence networks to foster the emergence of meaning. In International Conference on Complex Networks and Their Applications (pp. 377-388).
A complete SINr implementation can be found here : https://github.com/SINr-Embeddings/sinr But the code we provide in this pull request is enough to train efficiently node embeddings using networkx and scipy. We would really appreciate our method to be included in the karateclub package that we use a lot !