inria-soda/sepal-datasets

SEPAL: Scalable Feature Learning on Huge Knowledge Graphs for Downstream Machine Learning

This dataset contains the Mini YAGO3 knowledge graph and the downstream tasks used in the SEPAL paper (https://arxiv.org/pdf/2507.00965v2).

SEPAL is a knowledge graph embedding method for very large knowledge graphs. It is designed to produce good embeddings for downstream regression and classification tasks.

The code for SEPAL can be found at https://github.com/soda-inria/sepal

Descriptions of downstream tables

Real-world datasets

• US elections
  (https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/VOQCHQ)
  Voting statistics in the 2020 US presidential election. The task is to predict the number of votes per party in US counties.
• Housing prices
  (https://www.zillow.com/research/data)
  The task is to predict the average housing price in US cities.
• US accidents
  (https://smoosavi.org/datasets/us_accidents)
  Information on accidents in US cities between 2016 and 2020. The task is to predict the number of accidents in US cities.
• Movie revenues
  (https://www.kaggle.com/rounakbanik/the-movies-dataset)
  Metadata of movies released before July 2017. The task is to predict the box-office revenues.

WikiDBs datasets

42 datasets selected and curated from the WikiDBs [1] database (26 for classification, 16 for regression). Details in the paper.

The Mini YAGO3 knowledge graph

The Mini YAGO3 dataset is constructed from YAGO3 [2] by first filtering for entities with a degree of 9 or greater, and then selecting the largest connected component from the resulting subgraph. It contains 129,493 entities, 1,132,010 triples, and 37 distinct relations.

References

[1] Liane Vogel, Jan-Micha Bodensohn, and Carsten Binnig. "Wikidbs: A large-scale corpus of relational databases from wikidata". In Advances in Neural Information Processing Systems, 37:41186–41201, 2024.

[2] Farzaneh Mahdisoltani, Joanna Biega, and Fabian Suchanek. "Yago3: A knowledge base from multilingual wikipedias". In 7th biennial conference on innovative data systems research. CIDR Conference, 2014.