计算机科学与探索2026，Vol.20Issue(9)：2602-2611,10.DOI:10.3778/j.issn.1673-9418.2510055
面向时序知识图谱表示学习的统一评估基准
Unified Evaluation Benchmark for Temporal Knowledge Graph Representation Learning
摘要
Abstract
Temporal knowledge graph representation learning(TKGR-LE)aims to embed dynamic facts and entity rela-tionships into a low-dimensional vector space to capture the semantic evolution along the time dimension.This technique is widely applied in various scenarios,such as financial risk control,recommender systems,and event prediction.However,despite its promising outlook,the development of this field is significantly hampered by a profound lack of standardiza-tion.Existing studies often focus exclusively on the construction and validation of a single model,resulting in a deficit of a unified,reproducible experimental framework and a set of objective,standardized evaluation protocols.This"siloed"approach severely limits the performance comparison and technical exchange among different models,greatly constraining the overall progress and rapid iteration of algorithms in this domain.To address this critical bottleneck,this paper proposes a unified framework for temporal knowledge graph representation learning.The core objective of this framework is to establish a standardized system for research,development,and evaluation.Utilizing a modular design paradigm,it stan-dardizes the entire workflow,encompassing model design,implementation,and evaluation,and constructs an integrated experimental platform.Specifically,this paper systematically reproduces and optimizes eight categories of temporal algo-rithms,including TTransE and TA-DistMult.This paper conducts temporal-aware link prediction experiments on standard benchmark datasets,such as ICEWS and YAGO.Furthermore,multi-dimensional comparative experiments are performed to validate the advantages of this framework in algorithmic compatibility,evaluation objectivity,and scenario universality.This research endeavors to establish a standardized R&D and evaluation system for temporal knowledge graphs.By intro-ducing a normalized performance comparison mechanism,it provides a crucial baseline framework to facilitate future algorithmic iteration and technological innovation.关键词
时序知识图谱/表示学习/链接预测Key words
temporal knowledge graph/representation learning/link prediction分类
信息技术与安全科学引用本文复制引用
陈逸婧,李倩,杜云涛,刘磊,闫中敏..面向时序知识图谱表示学习的统一评估基准[J].计算机科学与探索,2026,20(9):2602-2611,10.基金项目
国家自然科学基金(62402293) （62402293）
国网山东省电力公司科技项目(520614250004) （520614250004）
山东省自然科学基金(ZR2025QC1570).This work was supported by the National Natural Science Foundation of China(62402293),the Science and Technology Project of State Grid Shandong Electric Power Company(520614250004),and the Natural Science Foundation of Shandong Province(ZR2025QC1570). （ZR2025QC1570）
