e The RDF2vec Family of Knowledge Graph Embedding Methods

The RDF2vec Family of Knowledge Graph Embedding Methods

Supplementary Material for the Semantic Web Journal Article

Introduction

This page collects information on how to reproduce the experiments in our Semantic Web Journal submission The RDF2vec Family of Knowledge Graph Embedding Methods.

The following sections contain pointers to repositories with the respective code, as well as short guidelines on how to reproduce the experiments.

Reproducing Results on the GEval Benchmark

The following steps need to be taken to reproduce the results on the GEval benchmark reported in section 7.2 of the paper:

  1. Download GEval from here.
  2. Download the pre-trained DBpedia vectors (.txt files) from here. The folder contains vectors both for all variants of RDF2vec used in the paper, as well as the seven link prediction embeddings used for comparison.
  3. Run the corresponding tasks on the GEval framework together with the .txt files downloaded in the previous step, following this recipe.

Reproducing Results on the DLCC DBpedia Benchmark

The following steps need to be taken to reproduce the results on the DBpedia based part of the DLCC benchmark reported in section 7.3 of the paper:

  1. Download DLCC Evaluation Module from here.
  2. Download the pre-trained DBpedia vectors (.txt files) from here. The folder contains vectors both for all variants of RDF2vec used in the paper, as well as the seven link prediction embeddings used for comparison.
  3. Use the EvaluationManager together with the .txt files downloaded in the previous step, following the minimal Python example here.

Reproducing Results on the DLCC Synthetic Benchmark

The following steps need to be taken to reproduce the results on the synthetic part of the DLCC benchmark reported in section 7.3 of the paper:

  1. Download DLCC Evaluation Module from here.
  2. Download the synthetic test cases from here. For each test case, there is a graph.nt file, as well as the lists of positives and negatives used for training and testing.
  3. Embed each graph.nt file to create embedding vectors, as well as pre-processed files for DGL-KE in the dgl-ke-graph directory. For RDF2vec, follow the minimal example here. For the baseline models, use DGL-KE, following the user-defined knowledge graph recipe here.
  4. With the test files and vectors obtained in the previous steps, use the EvaluationManager together with the .txt files downloaded in the previous step, following the minimal Python example here.