Generate embeddings on the graph, with RDF2Vec - using pyRDF2Vec -, TransE, DistMult - using pyKEEN.

In the config folder, edit appropriately the following files:

• object_properties.txt the list of object properties which will be taken into account for the random walks (1 per line);
• prefixes.txt to be added to the SPARQL queries
• get_entities.rq the query for getting the URIs of all entities of interest.

Run the following commands for downloading the data on your machine:

pip install -r requirements.txt
python preprocessing.py

Finally, generate the embeddings using:

python main.py [entities list] [-a algorith_name]

where entities list is a list of entities uri (1 per line) in a textual file. Following the generated files by the preprocessing, you can run (for example):

python main.py voc
python main.py smells -a TransE

This is producing an [entity].kv file, which is a gensim's KeyedVector file.

For generating a dense subgraph, you should run the following scripts

python reduce_graph.py
python generate_dense_subgraph.py

and run again the embedding script with

python main.py [entities list] [-a algorith_name] -d dense_graph.csv

Load embeddings in this way:

emb = KeyedVectors.load('emb.kv')

Search the most similar to a term:

emb.most_similar('http://data.odeuropa.eu/vocabulary/olfactory-objects/269', topn=10) # incense

# 0.7755   http://data.odeuropa.eu/vocabulary/olfactory-objects/267   Frankincense

Refer to gensim's documentation for further possibilities.

We performed experiment for clustering and link predicting using the following code:

• Clustering: notebook - [notebook with dense graph](./Clustering-Dense Graph.ipynb)
• Link prediction with TransE and DistMult: script