pmonnin / gcn-matching

Matching nodes in a knowledge graph using Graph Convolutional Networks and investigating the interplay between formal semantics and GCNs

gcn-matching

Matching nodes in a knowledge graph using Graph Convolutional Networks and investigating the interplay between formal semantics and GCNs.

A detailed description of the motivation and the algorithms is available in the related paper.

Citing

When citing, please use the following reference:

Pierre Monnin, Chedy Raïssi, Amedeo Napoli, Adrien Coulet: Discovering alignment relations with Graph Convolutional Networks: A biomedical case study. Semantic Web 13(3): 379-398 (2022)

@article{monninRNC22,
  author    = {Pierre Monnin and
               Chedy Ra{\"{\i}}ssi and
               Amedeo Napoli and
               Adrien Coulet},
  title     = {Discovering alignment relations with Graph Convolutional Networks:
               {A} biomedical case study},
  journal   = {Semantic Web},
  volume    = {13},
  number    = {3},
  pages     = {379--398},
  year      = {2022},
  url       = {https://doi.org/10.3233/SW-210452},
  doi       = {10.3233/SW-210452}
}

Scripts

1. Query similarity set

• In query_simset.py
• Retrieve individuals to match (instances of classes in individuals-classes in the JSON configuration file)
• Retrieve similarity links between these individuals (to use in train/valid/test sets).
  • Similarity links are described in similarity-links in the JSON configuration file
  • When having the link (url1, url2), we do not add (url2, url1) for symmetric predicates to avoid the symmetry bias in training

2. Query graph

• In query_graph.py
• Retrieve the adjacency of the RDF graph (except similarity links previously retrieved in 1.)
• Retrieve predicates and their inverses (or symmetry)
• Must be used with the cache manager resulting from the previous step

3. Similarity analysis

• In similarity_analysis.py
• Output PDF files with histograms depicting the size of similarity clusters and number of them for each model (computed based on the similarity links considered by each model)
• Similarity clusters for each model are computed over all similarity links indifferently considered by the model in an undirected (symmetry) and transitive fashion

4. N Fold Split

• In n_fold_split.py
• Output a n-fold split of similarity links (after shuffling)

5. Transform graph

• In transform_graph.py
• Output a DGL graph from the given RDF graph applying one of the following transformations:
  • G₀: RDF graph + adding an abstract inverse for each predicate
  • G₁: RDF graph after owl:sameAs edges contraction (only considering canonical nodes)
  • G₂: RDF graph with consideration of inverse predicates / symmetry (to avoid adding abstract inverses when not needed)
  • G₃: RDF graph with links added based on the hierarchy of predicates: if (a, rel₁, b) and (rel₁, subPropertyOf, rel₂), we add (a, rel₂, b)
  • G₄: RDF graph with rdf:type links added based on the hierarchy of classes: if (a, type, b) and (b, subClassOf, c), we add (a, type, c)
  • G₅: all transformations of G₁ to G₄
• The graph is limited to the considered neighborhood of individuals to match based on the number of layers

6. Learning

• In learning.py
• Output a python dict where each key is the index of the test fold and contains:
  • logits_history: python list associating an epoch with its logits
  • train_loss_history: python list associating an epoch with its train loss
  • val_loss_history: python list associating an epoch with its validation loss
  • test_loss_history: python list associating an epoch with its test loss
  • temperature_history: python list associating an epoch with its temperature
  • model: python list associating an epoch with the parameters of the GCN model

7. Clustering analysis

• In clustering_analysis.py
• Output for each fold:
  • A distance analysis based on all links, links whose nodes are in the training set, in the validation set, and in the test set
  • A UMAP projection computed on all nodes and displayed for all nodes, nodes in the training set, in the validation set and in the test set. Only --umap-colors similarity clusters are colored (starting at the biggest ones). Only similarity clusters containing more than --umap-size nodes are displayed (0 to disable)
  • A plot of the training, validation, and test losses
  • A plot of the temperature

Dependencies

• Python3.7
• tqdm
• requests
• pytorch
• dgl
• matplotlib
• scikit-learn
• umap-learn
• pynndescent

Experiments

Models

(called gold clusterings in the preprint)

Similarity links	owl:sameAs	skos:closeMatch	skos:relatedMatch	skos:related	skos:broadMatch
Properties	T / S	T / S	nT / S	nT / S	T / nS
M0	X	X	X	X	X
M1	X	X	X	X
M2	X
M3		X
M4			X
M5				X
M6					X

• T: transitivity
• S: symmetry
• nT : non-transitivity
• nS : non-symmetry