This code is part of a larger work that compares typical knowledge graph embeddings for Data Mining (e.g. RDF2vec) with those for Link Prediction (e.g. TransE). A version of the paper that is currently being reviewed is available here.
Here, we compare embedding approaches that are orignally designed for Data Mining (RDF2vec, Node2vec, DeepWalk) with Link Prediction approaches (TransE, TransR, RotatE, DistMult, RESCAL, ComplEx) on the following Data Mining tasks:
- Classification
- Clustering
- Regression
- Semantic Analogies
- Document Similarity
- Entity Relatedness
All these tasks are based on entities of the DBpedia 2016-10 dataset. Consequently, we use this dataset to create our embedding vectors.
To produce the embedding vectors and evaluate them on the mentioned Data Mining tasks, we use the following frameworks:
- KGvec2go to retrieve embedding vectors for RDF2vec
- DGL-KE to train embedding vectors for Link Prediction approaches
- GEval to evaluate the generated embedding vectors on Data Mining tasks
- Clone the GEval repository:
git clone [email protected]:mariaangelapellegrino/Evaluation-Framework.git evaluation_framework - Install the requirements:
pip install -r evaluation_framework/requirements.txt(you may want to use a virtual environment)
Download the embedding vectors for all approaches (RDF2vec, Node2vec, DeepWalk, KGlove, RDF2vecoa TransE, TransR, RotatE, DistMult, RESCAL, ComplEx) here and put them into the data folder.
Alternatively, you can use this notebook to train the embedding vectors of the link prediction approaches on your own. But note that it takes multiple hours on a GPU to train embedding vectors of a single approach due to the large size of the DBpedia dataset. So you have to plan 3-4 days for the generation of all embedding vectors.
To run the evaluation framework with the embedding approaches, use this notebook. If you are using virtual environments, make sure that the notebook runs in the same environment where the dependencies of GEval are installed.