KGE and GNN - support and examples

[ceteri]

Add support for knowledge graph embedding (KGE), to build on the UMAP example, based on DeepWalk, node2vec, etc., used to handle the semantic tagging of the full recipe dataset.

Depends on: #29

Tutorials to evaluate:

• https://towardsdatascience.com/node2vec-embeddings-for-graph-data-32a866340fef
• https://medium.com/@st3llasia/graph-embedding-techniques-7d5386c88c5
• https://maelfabien.github.io/machinelearning/graph_5/
• https://dmccreary.medium.com/understanding-graph-embeddings-79342921a97f
• https://towardsdatascience.com/how-to-create-representations-of-entities-in-a-knowledge-graph-using-pyrdf2vec-82e44dad1a0

Code repos to evaluate:

• https://github.com/eliorc/node2vec
• https://github.com/phanein/deepwalk
• https://github.com/benedekrozemberczki/graph2vec
• https://github.com/palash1992/GEM
• https://github.com/shenweichen/GraphEmbedding
• https://github.com/feixyz10/network-representation-learning
• https://github.com/IBCNServices/pyRDF2Vec
• https://github.com/THUMNLab/AutoGL

Papers to recommend:

• "Machine Learning on Graphs: A Model and Comprehensive Taxonomy"; Ines Chami, et al. (2020-05-07)
• PapersWithCode
• "Semantic Preserving Embeddings for Generalized Graphs"

[ceteri]

Tutorial about PyTorch Geometric https://towardsdatascience.com/hands-on-graph-neural-networks-with-pytorch-pytorch-geometric-359487e221a8

[dvsrepo]

Hi Paco!

I will start working on this next week, PyTorch geometric seems like a good choice to start working out the details, focusing on an application to the full recipe example.

[ceteri]

@dvsrepo just added as_tensor() method to SubgraphTensor to support the PyTorch_Geometric use case with Rubrix

[igrangel]

@ceteri @dvsrepo do you have an example of the integration with PyTorch Geometric? Currently, I'm using the following:

from torch_geometric.utils import from_networkx pG = from_networkx(G)

Assuming that the G is the result of a SPARQL query fetching everything in the graph. This is for sure not a good solution, I would be interested in your solution here.

[dvsrepo]

Hi Irlan! Nice to hear from you 😊

The current implementation is quite slow and I've planning to improve it for a long time now :( hopefully will do this beginning of September, but it uses the Subgraph class, you can find the reference here: https://derwen.ai/docs/kgl/ref/#kglab.SubgraphTensor.as_tensor

A faster way to do it is shown in the following tutorial which also shows how to train a gcn and transform URIs to tensor IDs and viceversa: https://docs.rubrix.ml/en/stable/tutorials/03-kglab_pytorch_geometric.html#2.-Representing-our-knowledge-graph-as-a-PyTorch-Tensor

Any suggestions to improve the implementation in kglab is more than welcome