DBpedia RDF2Vec Graph Embeddings

DBpedia graph embeddings using RDF2Vec. RDF2Vec embedding generation code can be found here and is based on a publication by Portisch et al. [1].

The embeddings dataset consists of 200-dimensional vectors of DBpedia entities (from 1/9/2021).

Generating Embeddings

The code for generating these embeddings can be found here.

Run the run.sh script that wraps all the necessary commmands to generate embeddings

bash run.sh

The script downloads a set of DBpedia files, which are listed in dbpedia_files.txt. It then builds a Docker image and runs a container of that image that generates the embeddings for the DBpedia graph defined by the DBpedia files.

A folder files is created containing all the downloaded DBpedia files, and a folder embeddings/dbpedia is created containing the embeddings in vectors.txt along a set of random walk files.

Run Time of Embeddings Generation

Generating embeddings can take more than a day, but it depends on the number of DBpedia files chosen to be downloaded. Following are some basic run time statistics when embeddings are generated on a 64 GB RAM, 8 cores (AMD EPYC), 1 TB SSD, 1996.221 MHz machine.

Total: 1 day, 8 hours, 52 minutes, 41 seconds

Walk generation: 0 days, 7 minutes, 24 minutes, 36 seconds

Training: 1 day, 1 hour, 28 minutes, 5 seconds

Parameters Used

Here is listed the parameters used to generate the embeddings provided here:

Number of walks per entity: 100

Depth (hops) per walk: 4

Walk generation mode: RANDOM_WALKS_DUPLICATE_FREE

Threads: # of processors / 2

Training mode: sg

Embeddings vector dimension: 200

Minimum word2vec word count: 1

Sample rate: 0.0

Training window size: 5

Training epochs: 5