| 引用本文： |
-
陈恺,朱东,李楠,于宿洋,李爱平.基于信息瓶颈的统一时序知识图谱推理模型[J].信息安全学报,已采用 [点击复制]
- chenkai,zhudong,linan,yusuyang,liaiping.Information Bottleneck-based Unified Reasoning for Temporal Knowledge Graphs[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 时序知识图谱推理中内插与外推任务相互割裂，对于需要兼顾历史补全与未来预测的任务，当前方法通常无法共享同一套表征体系，直接融合历史与未来信息会引入大量关系冗余与语义冲突，严重制约了时序知识图谱在动态、连续决策场景中的实用价值。本文提出一种统一的时序路径信息瓶颈推理模型（Temporal Path Reasoning model with Information Bottleneck，TPRIB），该模型在统一的查询中心子图提取、信息瓶颈路径压缩、神经驱动符号推理与联合优化框架下，分别适配内插与外推任务的信息可见性约束：内插任务允许在已知历史时间范围内利用双向时间邻域信息，外推任务则严格限制为查询时刻之前的历史事实，从而实现内插和外推的统一框架。对于大量关系冗余和语义冲突，TPRIB将信息瓶颈理论融入神经符号推理，以自动化提取推理过程中的关键时序路径并实现稳健高效推理。在多个基准数据集上的实验结果表明，所提TPRIB模型在内插与外推设定下的链接预测任务中取得了具有竞争力的性能，并展现出较好的可解释性与鲁棒性。此外，通过详细的案例分析与路径分析，我们发现TPRIB的输出不易受短路径或表面关联的干扰，能够稳健地进行长距离、多步的时序逻辑推理，有效应对数据中的噪声与不确定性。本文所提出的统一推理框架缓解了内插与外推任务割裂的难题，通过信息瓶颈理论与神经符号推理的有机融合，为时序知识的协同感知、补全与预测提供了可靠且高效的解决方案。 |
| 关键词: 时序知识图谱 信息瓶颈 知识推理 神经符号推理 |
| DOI： |
| 投稿时间：2026-03-11修订日期：2026-07-15 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Information Bottleneck-based Unified Reasoning for Temporal Knowledge Graphs |
|
chenkai, zhudong, linan, yusuyang, liaiping
|
| (National University of Defence Technology) |
| Abstract: |
| In temporal knowledge graph reasoning, the tasks of interpolation and extrapolation are often treated separately. For appli-cations that require both historical completion and future prediction, existing methods fail to share a unified representation framework. Directly combining historical and future information often introduces substantial relational redundancy and semantic conflicts, which severely limits the practical applicability of temporal knowledge graphs in dynamic and continuous decision-making scenarios. To address this issue, we propose a unified Temporal Path Reasoning model with Information Bottleneck (TPRIB). The model adapts to the information visibility constraints of interpolation and extrapolation tasks, respectively, under a unified framework integrating query-centric subgraph extraction, information bottleneck–based path compression, neurally driven symbolic reasoning, and joint optimization: interpolation tasks permit the utilization of bi-directional temporal neighborhood information within known historical time ranges, while extrapolation tasks are strictly limited to historical facts prior to the query moment, thereby achieving a unified framework for both interpolation and extrapolation. To mitigate relational redundancy and semantic conflicts, TPRIB incorporates information bottleneck theory into a neu-ro-symbolic reasoning framework, enabling the automatic identification of critical temporal reasoning paths and facilitating robust and efficient inference. Experimental results on multiple benchmark datasets demonstrate that TPRIB achieves competitive or superior performance on most link prediction settings under interpolation and extrapolation tasks, while exhibiting good interpretability and robustness. Furthermore, through detailed case studies and path-level analyses, we observe that TPRIB’s predictions are less susceptible to misleading short paths or superficial correlations. Instead, the model is capable of performing stable long-range, multi-step temporal logical reasoning, effectively handling noise and uncertainty in the data. Overall, the unified reasoning framework proposed in this work provides a methodological solution to the long-standing separation between interpolation and extrapolation tasks. By organically integrating information bottleneck theory with neuro-symbolic reasoning, it offers a reliable and efficient approach for the collaborative perception, completion, and pre-diction of temporal knowledge. |
| Key words: Temporal Knowledge Graph, Information Bottleneck, Knowledge Reasoning, Neural-Symbolic Reasoning |
