文章背景与核心概要

在持续知识图谱嵌入（Continual Knowledge Graph Embedding, CKGE）领域，模型必须随着图谱的增长不断更新实体和关系的表征。虽然现有的研究主要集中于解决“灾难性遗忘”（catastrophic forgetting），但本文指出了一个独特且常被忽视的现象：候选集干扰（Candidate-Set Interference）。

随着新实体被添加到知识图中，任意给定查询的“候选全集”会随之扩大。这导致历史答案的相对排名下降，即便它们在旧实体当中的原始得分和排序保持不变。为了解决这一问题，作者引入了匹配超额超越正则化（Matched Excess-Outranker Regularization, MEOR）。MEOR 作为一种宿主级别的目标函数，仅在来自新实体的竞争超过“匹配参考”（matched reference）时施加单侧惩罚，从而在无需改变模型架构的情况下有效保留历史排名。

Matched Excess-Outranker Regularization for Candidate-Set Interference in Continual Knowledge Graph Embedding

Authors: Hao Ren, Junbin Gao, Jiaojiao Jiang
Date: August 25, 2026
Subject: Artificial Intelligence (cs.AI); Databases (cs.DB); Information Retrieval (cs.IR)
arXiv ID: 2608.24273

摘要

In the context of Continual Knowledge Graph Embedding (CKGE), models must update entity and relation representations as graphs grow. While existing research focuses on "catastrophic forgetting," this paper identifies a distinct, overlooked phenomenon: Candidate-Set Interference.

As new entities are added to a knowledge graph, the "candidate universe" for any given query expands. This causes historical answers to lose their relative rank, even if their original scores and ordering among old entities remain preserved. To address this, the authors introduce Matched Excess-Outranker Regularization (MEOR). MEOR acts as a host-level objective that applies a one-sided penalty only when competition from new entities exceeds a "matched reference," effectively preserving historical ranking without requiring architectural changes.

核心贡献

• 候选集干扰的形式化定义： 作者定义了实体集的扩展如何本质上导致历史答案排名的下降，并将其与传统的灾难性遗忘区分开来。
• 引入 MEOR： 一种新颖的正则化技术，它将平滑的答案相关新来者压力与对分数盲目、结构匹配的旧参考进行对比。
• 性能提升：
  • 在 ENTITY-ComplEx 的八次配对运行中，与标准重放方法相比，MEOR 将历史当前全集平均倒数排名（MRR）提高了 0.0057。
  • 将候选集干扰降低了 0.0055。
  • 在所有十个测试的 FBInc-S 和 FBInc-L 设置中展现出一致的性能提升。
• 兼容性： MEOR 被证明是一个即插即用的解决方案，能够在无需替换底层嵌入架构或持续学习器的情况下提升性能。

Key Contributions

• Formalization of Candidate-Set Interference: The authors define how the expansion of the entity set inherently degrades the ranking of historical answers, distinguishing this from traditional catastrophic forgetting.
• Introduction of MEOR: A novel regularization technique that compares smooth answer-relative newcomer pressure against score-blind, structurally matched old references.
• Performance Gains:
  • Across eight paired runs on ENTITY-ComplEx, MEOR improved historical current-universe Mean Reciprocal Rank (MRR) by 0.0057 over standard replay methods.
  • Reduced candidate-set interference by 0.0055.
  • Demonstrated consistent improvements across all ten tested FBInc-S and FBInc-L settings.
• Compatibility: MEOR is shown to be a plug-and-play solution that improves performance without necessitating the replacement of the underlying embedding architecture or the continual learner.

技术元数据

Technical Metadata

Attribute	Details
Cite as	arXiv:2608.24273 [cs.AI]
DOI	10.48550/arXiv.2608.24273
Full-text	PDF | HTML | TeX Source

如需进一步探索，包括文献计量工具和相关论文，请访问 arXiv 官方页面。

For further exploration, including bibliographic tools and related papers, please visit the official arXiv page.