Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Breaking the Reasoning Horizon in Entity Alignment Foundation Models

Yuanning Cui, Zequn Sun, Wei Hu, Kexuan Xin, Zhangjie Fu

TL;DR

This work tackles the transferability challenge in entity alignment by identifying a reasoning horizon gap that arises when adapting graph foundation models to cross-KG EA. It introduces EAFM, an EA foundation model with parallel encoding guided by seed EA anchors and a global merged relation graph, plus a learnable interaction module and a bidirectional objective to achieve robust zero-shot transfer to unseen KGs. Substantial empirical results show EAFM outperforms state-of-the-art baselines across OpenEA, SRPRS, and DBpedia datasets, with strong inductive generalization and ablations confirming the importance of anchor-conditioned encoding and relation graph unification. The approach enables efficient, structure-driven EA that generalizes across heterogeneous schemas, offering a practical pathway toward universal, service-oriented EA for KG fusion.

Abstract

Entity alignment (EA) is critical for knowledge graph (KG) fusion. Existing EA models lack transferability and are incapable of aligning unseen KGs without retraining. While using graph foundation models (GFMs) offer a solution, we find that directly adapting GFMs to EA remains largely ineffective. This stems from a critical "reasoning horizon gap": unlike link prediction in GFMs, EA necessitates capturing long-range dependencies across sparse and heterogeneous KG structuresTo address this challenge, we propose a EA foundation model driven by a parallel encoding strategy. We utilize seed EA pairs as local anchors to guide the information flow, initializing and encoding two parallel streams simultaneously. This facilitates anchor-conditioned message passing and significantly shortens the inference trajectory by leveraging local structural proximity instead of global search. Additionally, we incorporate a merged relation graph to model global dependencies and a learnable interaction module for precise matching. Extensive experiments verify the effectiveness of our framework, highlighting its strong generalizability to unseen KGs.

Breaking the Reasoning Horizon in Entity Alignment Foundation Models

TL;DR

This work tackles the transferability challenge in entity alignment by identifying a reasoning horizon gap that arises when adapting graph foundation models to cross-KG EA. It introduces EAFM, an EA foundation model with parallel encoding guided by seed EA anchors and a global merged relation graph, plus a learnable interaction module and a bidirectional objective to achieve robust zero-shot transfer to unseen KGs. Substantial empirical results show EAFM outperforms state-of-the-art baselines across OpenEA, SRPRS, and DBpedia datasets, with strong inductive generalization and ablations confirming the importance of anchor-conditioned encoding and relation graph unification. The approach enables efficient, structure-driven EA that generalizes across heterogeneous schemas, offering a practical pathway toward universal, service-oriented EA for KG fusion.

Abstract

Entity alignment (EA) is critical for knowledge graph (KG) fusion. Existing EA models lack transferability and are incapable of aligning unseen KGs without retraining. While using graph foundation models (GFMs) offer a solution, we find that directly adapting GFMs to EA remains largely ineffective. This stems from a critical "reasoning horizon gap": unlike link prediction in GFMs, EA necessitates capturing long-range dependencies across sparse and heterogeneous KG structuresTo address this challenge, we propose a EA foundation model driven by a parallel encoding strategy. We utilize seed EA pairs as local anchors to guide the information flow, initializing and encoding two parallel streams simultaneously. This facilitates anchor-conditioned message passing and significantly shortens the inference trajectory by leveraging local structural proximity instead of global search. Additionally, we incorporate a merged relation graph to model global dependencies and a learnable interaction module for precise matching. Extensive experiments verify the effectiveness of our framework, highlighting its strong generalizability to unseen KGs.
Paper Structure (29 sections, 11 equations, 4 figures, 6 tables)

This paper contains 29 sections, 11 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work and Discussions
  3. Knowledge Graph Foundation Models
  4. Entity Alignment
  5. Methodology
  6. Problem Formulation
  7. Framework Overview
  8. Global Relation Learning via Merged Graph
  9. Relation Graph Construction.
  10. Structure-Aware RelGNN.
  11. Parallel Entity Encoding
  12. Anchor-Conditioned Initialization.
  13. Conditional Message Passing.
  14. Shortening Inference Trajectory.
  15. Interaction and Alignment Learning
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Illustration of the critical challenges in adapting GFMs to EA. (a) Reasoning horizon gap: The significant performance degradation with respect to reasoning distance empirically validates the difficulty of long-range inference in EA. (b) Semantic complexity gap: A comparison of the topological patterns between standard relations and the sameAs relation, illustrating the latter's significantly higher-order connectivity and complexity.
  • Figure 2: The overall architecture of the proposed entity alignment foundation model (PEA). (a) Structural Pre-training Phase: The model employs a parallel encoding strategy on source KGs. It jointly optimizes the RelGNN (via the Merge Relation Graph) and the EntGNN (via anchor-conditioned message passing) to capture transferable structural patterns. (b) Inductive Inference Phase: When transferring to unseen KGs, the pre-trained RelGNN and EntGNN parameters are frozen. The model directly performs alignment by initializing features based on local anchors in the new graphs, without requiring re-training or fine-tuning.
  • Figure 3: Impact of pre-training data characteristics on zero-shot transfer performance. The yellow bars consistently represent the anchor source D-W-15K-V1. The comparative bars illustrate the impact of specific variants: (a) the dense source D-W-15K-V2 (blue), (b) the large-scale source D-W-100K-V1 (red), and (c) the cross-lingual source EN-DE-15K-V1 (green).
  • Figure 4: Impact of the anchor-hop distance $k$ on MRR performance across different benchmark groups.