Python implementation and extension of RDF2Vec to create a 2D feature matrix from a Knowledge Graph for downstream ML tasks.
RDF2Vec is an unsupervised technique that builds further on Word2Vec, where an embedding is learned per word, in two ways:
- the word based on its context: Continuous Bag-of-Words (CBOW);
- 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]).
Recently, a book about RDF2Vec was published by Heiko Paulheim, Jan Portisch, and Petar Ristoski. The book is a great introduction to what RDF2Vec is, and what can be done with it. The examples in the book use pyRDF2Vec, so it is recommended to have a look at it!
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).
pyRDF2Vec can be installed in three ways:
- from PyPI using
pip:
pip install pyRDF2vec- from any compatible Python dependency manager (e.g.,
poetry):
poetry add pyRDF2vec- from source:
git clone https://github.com/IBCNServices/pyRDF2Vec.git
pip install .To create embeddings for a list of entities, there are two steps to do beforehand:
- use a KG;
- 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.
To use a KG, you can initialize it in three ways:
- 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"],
],
),- 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",
],
],
),- 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)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())]The extraction of walks can take hours, days if not more in some cases. That's
why it is important to use certain attributes and optimize pyRDF2Vec
parameters as much as possible according to your use cases.
This section aims to help you to set up these parameters with some advice.
By default multiprocessing is disabled (n_jobs=1). If your machine allows
it, it is recommended to use multiprocessing by incrementing the number of
processors used for the extraction of walks:
from pyrdf2vec.walkers import RandomWalker
RDF2VecTransformer(walkers=[RandomWalker(4, 10, n_jobs=4)])In the above snippet, the random walking strategy will use 4 processors to extract the walks, whether for a local or remote KG.
WARNING: using a large number of processors may violate the policy of some SPARQL endpoint servers. This being that using multiprocessing means that each processor will send a SPARQL request to one server to fetch the hops of the entity it is processing. Therefore, since these requests may take place in a short time, this server could consider them as a Denial-Of-Service attack (DOS). Of course, these risks are multiplied in the absence of cache and when the entities to be treated are of a consequent number.
By default the SPARQL requests bundling is disabled
(mul_req=False). However, if you are using a remote KG and have a large
number of entities, this option can greatly speed up the extraction of walks:
import pandas as pd
from pyrdf2vec import RDF2VecTransformer
from pyrdf2vec.graphs import KG
from pyrdf2vec.walkers import RandomWalker
data = pd.read_csv("samples/countries-cities/entities.tsv", sep="\t")
RDF2VecTransformer(walkers=[RandomWalker(4, 10)]).fit_transform(
KG("https://dbpedia.org/sparql", mul_req=True),
[entity for entity in data["location"]],
)In the above snippet, the KG specifies to the internal connector that it uses, to fetch the hops of the specified entities in an asynchronous way. These hops will then be stored in cache and be accessed by the walking strategy to accelerate the extraction of walks for these entities.
WARNING: bundling SPARQL requests for a number of entities considered too
large can may violate the policy of some SPARQL endpoint servers. As for the
use of multiprocessing (which can be combined with mul_req), sending a
large number of SPARQL requests simultaneously could be seen by a server as a
DOS. Be aware that the number of entities you have in your file corresponds to
the number of simultaneous requests that will be made and stored in cache.
By default, pyRDF2Vec uses a cache that provides a Least Recently Used
(LRU) policy, with a
size that can hold 1024 entries, and a Time To Live (TTL) of 1200 seconds. For
some use cases, you would probably want to modify the cache policy, increase (or decrease) the
cache size and/or change the TTL:
import pandas as pd
from cachetools import MRUCache
from pyrdf2vec import RDF2VecTransformer
from pyrdf2vec.graphs import KG
from pyrdf2vec.walkers import RandomWalker
data = pd.read_csv("samples/countries-cities/entities.tsv", sep="\t")
RDF2VecTransformer(walkers=[RandomWalker(4, 10)]).fit_transform(
KG("https://dbpedia.org/sparql", cache=MRUCache(maxsize=2048),
[entity for entity in data["location"]],
)By default, pyRDF2Vec uses [RandomWalker(2, None, UniformSampler())] as
walking strategy. Using a greater maximum depth indicates a longer extraction
time for walks. Add to this that using max_walks=None, extracts more walks
and is faster in most cases than when giving a number (SEE: FAQ).
In some cases, using another sampling strategy can speed up the extraction of walks by assigning a higher weight to some paths than others:
import pandas as pd
from pyrdf2vec import RDF2VecTransformer
from pyrdf2vec.graphs import KG
from pyrdf2vec.samplers import PageRankSampler
from pyrdf2vec.walkers import RandomWalker
data = pd.read_csv("samples/countries-cities/entities.tsv", sep="\t")
RDF2VecTransformer(
walkers=[RandomWalker(2, None, PageRankSampler())]
).fit_transform(
KG("https://dbpedia.org/sparql"),
[entity for entity in data["location"]],
)Loading large RDF files into memory will cause memory issues. Remote KGs serve as a solution for larger KGs, but using a public endpoint will be slower due to overhead caused by HTTP requests. For that reason, it is better to set up your own local server and use that for your "Remote" KG.
To set up such a server, a tutorial has been made on our wiki.
For more information on how to use pyRDF2Vec, visit our online documentation which is automatically updated
with the latest version of the main branch.
From then on, you will be able to learn more about the use of the modules as well as their functions available to you.
Your help in the development of pyRDF2Vec is more than welcome.
The architecture of pyRDF2Vec makes it easy to create new extraction and
sampling strategies, new embedding techniques. In order to better understand
how you can help either through pull requests and/or issues, please take a look
at the CONTRIBUTING
file.
pyRDF2Vec's walking strategies, sampling strategies and Word2Vec work with
randomness. To get reproducible embeddings, you firstly need to use a seed to
ensure determinism:
PYTHONHASHSEED=42 python foo.pyAdded to this, you must also specify a random state to the walking strategy which will implicitly use it for the sampling strategy:
from pyrdf2vec.walkers import RandomWalker
RandomWalker(2, None, random_state=42)NOTE: the PYTHONHASHSEED (e.g., 42) is to ensure determinism.
Finally, to ensure random determinism for Word2Vec, you must specify a single worker:
from pyrdf2vec.embedders import Word2Vec
Word2Vec(workers=1)NOTE: using the n_jobs and mul_req parameters does not affect the
random determinism.
Currently, the BFS function (using the Breadth-first search algorithm) is used
when max_walks=None which is significantly faster than the DFS function
(using the Depth-first search algorithm) and extract more walks.
We hope that this algorithmic complexity issue will be solved for the next
release of pyRDf2Vec
Sets the TCMALLOC_LARGE_ALLOC_REPORT_THRESHOLD environment variable to a
high value.
If you use pyRDF2Vec in a scholarly article, we would appreciate a
citation:
@inproceedings{pyrdf2vec,
title = {pyRDF2Vec: A Python Implementation and Extension of RDF2Vec},
author = {Steenwinckel, Bram and Vandewiele, Gilles and Agozzino, Terencio and Ongenae, Femke},
year = 2023,
publisher = {Springer Nature Switzerland},
booktitle = {European Semantic Web Conference},
doi = {10.1007/978-3-031-33455-9_28},
url = {https://arxiv.org/abs/2205.02283},
pages = {471--483},
}