Your Knowledge Graph Embeddings are Secretly Circuits and You Should Treat Them as SuchDownload PDF

Published: 26 Jul 2022, Last Modified: 17 May 2023TPM 2022Readers: Everyone
Keywords: Probabilistic Circuits, Knowledge Graph Embeddings, Tractable Inference
TL;DR: Tractable probabilistic Knowledge Graph Embeddings models and principled probabilistic complex query answering
Abstract: Some of the most popular and successful knowledge graph embedding (KGE) models---CP, ComplEx, RESCAL and TuckER---encode tensor factorizations that define an energy-based score over subject-relation-object triples. As such, they are not amenable to efficient maximum-likelihood training, and do not easily allow to sample triples nor answering complex queries in a principled probabilistic way. In this paper, we show how all these models can be readily interpreted as constrained computational graphs---circuits---and show how, by some minor modifications, one can turn them into tractable generative models of triples. This novel perspective not only fixes many of the aforementioned shortcomings of KGE models, but helps understand why recent learning strategies for KGE are successful while suggesting interesting new ones.
1 Reply

Loading