References: Knowledge Graph Foundation Models¶
-
Knowledge Graph - Wikipedia - Overview of knowledge graphs as structured representations of facts, covering history from Google's Knowledge Graph (2012) through current large-scale deployments; provides the definitional grounding for what "knowledge graph foundation models" are generalizing over.
-
Link Prediction - Wikipedia - Describes the link prediction task in networks, including transductive versus inductive settings; directly relevant to understanding why zero-shot link prediction across unseen knowledge graphs is a meaningful capability.
-
Inductive Transfer Learning - Wikipedia - Explains transfer learning paradigms, including the distinction between transductive and inductive transfer; grounds the conceptual leap from graph-specific trained models to universal pre-trained models like ULTRA.
-
Hamilton, W. L. — Graph Representation Learning — Morgan & Claypool, 2020 — The standard graduate-level textbook on graph ML; Chapter 7 covers knowledge graph embeddings and sets up the transductive framing that inductive and universal models (ULTRA, InGram) are designed to overcome.
-
Hogan, A., Blomqvist, E., Cochez, M., et al. — Knowledge Graphs — Synthesis Lectures on Data, Semantics, and Knowledge, Morgan & Claypool, 2021 — Comprehensive treatment of knowledge graph construction, representation, and reasoning; Chapter 6 on embeddings and Chapter 7 on rule mining provide context for why purely embedding-based approaches have scalability and generalization limits.
-
ULTRA: Towards Foundation Models for Knowledge Graph Reasoning - arXiv - Galkin et al. (2023) introduce ULTRA, showing that a single pre-trained model with double equivariance achieves competitive zero-shot link prediction across 43 knowledge graphs; the core reference for this chapter's universal graph model section.
-
NBFNet: Neural Bellman-Ford Networks for Link Prediction - arXiv - Zhu et al. (2021) present NBFNet, the path-based GNN backbone used inside ULTRA; understanding NBFNet's generalized Bellman-Ford message passing is essential for understanding how ULTRA achieves relation-conditioned entity scoring.
-
InGram: Inductive Knowledge Graph Embedding via Relation Graphs - arXiv - Lee et al. (2023) propose building a relation-level graph whose edges encode co-occurrence and inverse patterns, enabling embedding induction for both unseen entities and unseen relations; the key reference for the relation graph induction section.
-
ULTRA on Papers With Code — Knowledge Graph Completion Benchmarks - Papers With Code - Aggregates reported results for ULTRA on standard inductive KG benchmarks (FB15k-237, WN18RR, NELL-995 inductive splits); useful for locating updated leaderboard comparisons and reproduced numbers from the community.
-
PyTorch Geometric: Knowledge Graph Datasets and Link Prediction Utilities - PyTorch Geometric Docs - Documents the KG dataset classes (
FB15k237,WordNet18RR,OGB LinkPropPred) and link prediction loaders used to run ULTRA, NBFNet, and InGram experiments; the practical reference for reproducing this chapter's code examples.