entity2vec computes vector representations of Knowledge Graph entities that preserve semantic similarities and are suitable for classification tasks. It generates a set of property-specific entity embeddings by running node2vec on property specific subgraphs, i.e. K(p) = (s,p,o) where p is a given property. The repository includes:

• A reimplementation of node2vec, which introduces the possibility of avoiding the preprocessing of the transition probabilities, which has the effect of reducing memory effort, but slowing down the computation

• entity2vec, which generates a set of entity embeddings from Knowledge Graphs corresponding to different properties. Entity2vec can work with a set of pre-downloaded dumps or download them from a SPARQL endpoint.

• Python 2.7 or above
• numpy
• gensim
• networkx
• pandas
• SPARQL Wrapper

If you are using pip:

    pip install -r requirements.txt

The set of properties can be defined in the configuration file config/properties.json, otherwise the software will run on each file that is located in datasets/your_dataset/graphs or if a SPARQL endpoint is provided, it will download all the graphs for all properties in datasets/your_dataset/graphs.

    python src/entity2vec.py --dataset dataset --config_file config_file --entities entities --sparql sparql --default_graph default_graph

Alternatively, e2v can be loaded as a module and used like:

    from entity2vec.entity2vec import Entity2Vec

    e2v = Entity2Vec(False, False, False, 1, 1, 10, 5,
                 128, 10, 8, 5, 'path/to/properties.json', False,
                 'dataset_name', 'all', False, False,
                 'path/to/feedback.edgelist')

Generate unique vector representation for an entity, without considering the role of semantic properties, to use in classification tasks.

0. Create empty directory called emb

1. Run node2vec on the whole graph to create a single global embedding of the entity

    python src/node2vec.py --input datasets/aifb/aifb.edgelist --output emb/aifb_p1_q4.emd  --p 1 --q 4
   

2. Obtain scores, e.g.:

    cd ml
   
    python rdf_predict.py --dataset aifb --emb ../emb/aifb_p1_q4.emd --dimension 500