117 - Information-Theoretic Minimal Sufficient Representation for Multi-Source Knowledge Graph Completion
Association for Artificial Intelligence 2026, Weiyi Yang, Yongxiu Xu, Tingwen Liu, Jiawei Sheng, Taoyu Su, Linghui Wang
Open MIND
Problems Identified (5)
Redundant multi-domain KG representations: Existing multi-domain KG completion methods learn and fuse representations in ways that can be affected by redundant information within knowledge graphs.
Hidden task-relevant information: Redundant information can conceal task-relevant information in representations and impede improvements when scaling to many KGs.
Multi-source KG completion: Multi-domain knowledge graph completion aims to predict missing triples in a target KG using triples from multiple KGs in different domains.
Redundant multi-domain KG representations: Existing multi-domain KG completion methods learn and fuse representations in ways that can be affected by redundant information within knowledge graphs.
Hidden task-relevant information: Redundant information can conceal task-relevant information in representations and impede improvements when scaling to many KGs.
Proposed Solutions (5)
IMKGC minimal sufficient representation framework: IMKGC is an information-theoretic MKGC framework that learns minimal sufficient representations.
Context-complement-consistency preservation with variational suppression: IMKGC learns entity representations by preserving endogenous contextual, exogenous complementary, and equivalent-entity consistent information while suppressing redundant information via variational constraints.
Relation reasoning decoder: The framework uses a relation reasoning decoder to produce compressed relation representations and capture relatedness among relations.
IMKGC minimal sufficient representation framework: IMKGC is an information-theoretic MKGC framework that learns minimal sufficient representations.
Context-complement-consistency preservation with variational suppression: IMKGC learns entity representations by preserving endogenous contextual, exogenous complementary, and equivalent-entity consistent information while suppressing redundant information via variational constraints.
Results (3)
State-of-the-art MKGC performance:
Improved redundant-scenario performance:
Improved triple prediction:
Research Domain
Multi-domain knowledge graph completion / knowledge graphs