Hierarchically Metric-Structured Knowledge Graph Embeddings

TMLR Paper6448 Authors

Abstract: In the vast landscape of big data, there is an important challenge in understanding data and structuring it in a suitable format. Knowledge graphs are considered a sophisticated solution to organize and infer data and knowledge, offering a structured framework that transcends disciplinary boundaries in medicine, culture, biology, social networks, music, and beyond. Despite their informativeness, these systems are typically incomplete and their intrinsic structure unknown, whereas existing methodologies for predicting missing facts and characterizing their structure face scalability and interpretability issues. Addressing this gap, we introduce a new latent feature model, leveraging the prominent RESCAL framework to account for degree heterogeneity, multiscale structure, and scalable inference using an approximation of the full likelihood of all triplets circumventing negative sampling inference strategies. This not only enhances computational efficiency but also provides deeper insights into the intrinsic multiscale structure of knowledge graphs, thereby advancing the interpretability of predictive models and paving the way for a more comprehensive understanding of complex information networks.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=XDizCFJeBM&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)

Changes Since Last Submission: We updated the manuscript to use the latest TMLR template. The font size of the numbers in the tables was increased for improved readability, and all table captions were repositioned above the tables to comply with the template style. Additionally, the clustering results for the artificial knowledge graph and the learning curves were moved to the appendix.

Assigned Action Editor: ~Guillaume_Obozinski3

Submission Number: 6448