What is RDF2Vec?

RDF2Vec is an unsupervised technique that builds further on Word2Vec, where an embedding is learned per word, in two ways:

1. the word based on its context: Continuous Bag-of-Words (CBOW);

2. the context based on a word: Skip-Gram (SG).

To create this embedding, RDF2Vec first creates “sentences” which can be fed to Word2Vec by extracting walks of a certain depth from a Knowledge Graph.

This repository contains an implementation of the algorithm in “RDF2Vec: RDF Graph Embeddings and Their Applications” by Petar Ristoski, Jessica Rosati, Tommaso Di Noia, Renato De Leone, Heiko Paulheim ([paper] [original code]).

Getting Started

For most uses-cases, here is how pyRDF2Vec should be used to generate embeddings and get literals from a given Knowledge Graph (KG) and entities:

import pandas as pd

from pyrdf2vec import RDF2VecTransformer
from pyrdf2vec.embedders import Word2Vec
from pyrdf2vec.graphs import KG
from pyrdf2vec.walkers import RandomWalker

# Read a CSV file containing the entities we want to classify.
data = pd.read_csv("samples/countries-cities/entities.tsv", sep="\t")
entities = [entity for entity in data["location"]]
print(entities)
# [
#    "http://dbpedia.org/resource/Belgium",
#    "http://dbpedia.org/resource/France",
#    "http://dbpedia.org/resource/Germany",
# ]

# Define our knowledge graph (here: DBPedia SPARQL endpoint).
knowledge_graph = KG(
    "https://dbpedia.org/sparql",
    skip_predicates={"www.w3.org/1999/02/22-rdf-syntax-ns#type"},
    literals=[
        [
            "http://dbpedia.org/ontology/wikiPageWikiLink",
            "http://www.w3.org/2004/02/skos/core#prefLabel",
        ],
        ["http://dbpedia.org/ontology/humanDevelopmentIndex"],
    ],
)
# Create our transformer, setting the embedding & walking strategy.
transformer = RDF2VecTransformer(
    Word2Vec(epochs=10),
    walkers=[RandomWalker(4, 10, with_reverse=False, n_jobs=2)],
    # verbose=1
)
# Get our embeddings.
embeddings, literals = transformer.fit_transform(knowledge_graph, entities)
print(embeddings)
# [
#     array([ 1.5737595e-04,  1.1333118e-03, -2.9838676e-04,  ..., -5.3064007e-04,
#             4.3192197e-04,  1.4529384e-03], dtype=float32),
#     array([-5.9027621e-04,  6.1689125e-04, -1.1987977e-03,  ...,  1.1066757e-03,
#            -1.0603866e-05,  6.6087965e-04], dtype=float32),
#     array([ 7.9996325e-04,  7.2907173e-04, -1.9482171e-04,  ...,  5.6251377e-04,
#             4.1435464e-04,  1.4478950e-04], dtype=float32)
# ]

print(literals)
# [
#     [('1830 establishments in Belgium', 'States and territories established in 1830',
#       'Western European countries', ..., 'Member states of the Organisation
#       internationale de la Francophonie', 'Member states of the Union for the
#       Mediterranean', 'Member states of the United Nations'), 0.919],
#     [('Group of Eight nations', 'Southwestern European countries', '1792
#       establishments in Europe', ..., 'Member states of the Union for the
#       Mediterranean', 'Member states of the United Nations', 'Transcontinental
#       countries'), 0.891]
#     [('Germany', 'Group of Eight nations', 'Articles containing video clips', ...,
#       'Member states of the European Union', 'Member states of the Union for the
#       Mediterranean', 'Member states of the United Nations'), 0.939]
#  ]

If you are using a dataset other than MUTAG (where the interested entities have no parents in the KG), it is highly recommended to specify with_reverse=True (defaults to False) in the walking strategy (e.g., RandomWalker). Such a parameter allows Word2Vec to have a better learning window for an entity based on its parents and children and thus predict test data with better accuracy.

In a more concrete way, we provide a blog post with a tutorial on how to use pyRDF2Vec here.

NOTE: this blog uses an older version of pyRDF2Vec, some commands need be to adapted.

If you run the above snippet, you will not necessarily have the same embeddings, because there is no conservation of the random determinism, however it remains possible to do it (SEE: FAQ).

Installation

pyRDF2Vec can be installed in three ways:

1. from PyPI using pip:

pip install pyRDF2vec

2. from any compatible Python dependency manager (e.g., poetry):

poetry add pyRDF2vec

3. from source:

git clone https://github.com/IBCNServices/pyRDF2Vec.git
pip install .

Introduction

To create embeddings for a list of entities, there are two steps to do beforehand:

1. use a KG;

2. define a walking strategy.

For more elaborate examples, check the examples folder.

If no sampling strategy is defined, UniformSampler is used. Similarly for the embedding techniques, Word2Vec is used by default.

Use a Knowledge Graph

To use a KG, you can initialize it in three ways:

1. From a endpoint server using SPARQL:

from pyrdf2vec.graphs import KG

# Defined the DBpedia endpoint server, as well as a set of predicates to
# exclude from this KG and a list of predicate chains to fetch the literals.
KG(
    "https://dbpedia.org/sparql",
    skip_predicates={"www.w3.org/1999/02/22-rdf-syntax-ns#type"},
    literals=[
        [
            "http://dbpedia.org/ontology/wikiPageWikiLink",
            "http://www.w3.org/2004/02/skos/core#prefLabel",
        ],
        ["http://dbpedia.org/ontology/humanDevelopmentIndex"],
     ],
 ),

2. From a file using RDFLib:

from pyrdf2vec.graphs import KG

# Defined the MUTAG KG, as well as a set of predicates to exclude from
# this KG and a list of predicate chains to get the literals.
KG(
    "samples/mutag/mutag.owl",
    skip_predicates={"http://dl-learner.org/carcinogenesis#isMutagenic"},
    literals=[
        [
            "http://dl-learner.org/carcinogenesis#hasBond",
            "http://dl-learner.org/carcinogenesis#inBond",
        ],
        [
            "http://dl-learner.org/carcinogenesis#hasAtom",
            "http://dl-learner.org/carcinogenesis#charge",
        ],
    ],
),

3. From scratch:

from pyrdf2vec.graphs import KG, Vertex

 GRAPH = [
     ["Alice", "knows", "Bob"],
     ["Alice", "knows", "Dean"],
     ["Dean", "loves", "Alice"],
 ]
 URL = "http://pyRDF2Vec"
 CUSTOM_KG = KG()

 for row in GRAPH:
     subj = Vertex(f"{URL}#{row[0]}")
     obj = Vertex((f"{URL}#{row[2]}"))
     pred = Vertex((f"{URL}#{row[1]}"), predicate=True, vprev=subj, vnext=obj)
     CUSTOM_KG.add_walk(subj, pred, obj)

Define Walking Strategies With Their Sampling Strategy

All supported walking strategies can be found on the Wiki page.

As the number of walks grows exponentially in function of the depth, exhaustively extracting all walks quickly becomes infeasible for larger Knowledge Graphs. In order to avoid this issue, sampling strategies can be applied. These will extract a fixed maximum number of walks per entity and sampling the walks according to a certain metric.

For example, if one wants to extract a maximum of 10 walks of a maximum depth of 4 for each entity using the random walking strategy and Page Rank sampling strategy, the following code snippet can be used:

from pyrdf2vec.samplers import PageRankSampler
from pyrdf2vec.walkers import RandomWalker

walkers = [RandomWalker(4, 10, PageRankSampler())]