Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RLGNet: Repeating-Local-Global History Network for Temporal Knowledge Graph Reasoning

Ao Lv, Guige Ouyang, Yongzhong Huang, Yue Chen, Haoran Xie

TL;DR

RLGNet tackles the challenging problem of extrapolating future events in Temporal Knowledge Graphs by integrating multi-scale historical information through a three-module ensemble: Repeating History, Local History, and Global History. The Local module uses an RNN to capture short-term patterns, the Global module employs an MLP to model long-term trends, and the Repeating module reinforces recurring facts; together they produce a final score via a weighted combination with parameter $\alpha$. The approach is trained with independent losses per module and leverages candidate entity constraints to reduce noise. Empirical results on six benchmark datasets show consistent improvements in both single-step and multi-step reasoning, highlighting the benefit of a multi-scale ensemble design for temporal reasoning tasks.

Abstract

Temporal Knowledge Graph (TKG) reasoning involves predicting future events based on historical information. However, due to the unpredictability of future events, this task is highly challenging. To address this issue, we propose a multi-scale hybrid architecture model based on ensemble learning, called RLGNet (Repeating-Local-Global History Network). Inspired by the application of multi-scale information in other fields, we introduce the concept of multi-scale information into TKG reasoning. Specifically, RLGNet captures and integrates different levels of historical information by combining modules that process information at various scales. The model comprises three modules: the Repeating History Module focuses on identifying repetitive patterns and trends in historical data, the Local History Module captures short-term changes and details, and the Global History Module provides a macro perspective on long-term changes. Additionally, to address the limitations of previous single-architecture models in generalizing across single-step and multi-step reasoning tasks, we adopted architectures based on Recurrent Neural Networks (RNN) and Multi-Layer Perceptrons (MLP) for the Local and Global History Modules, respectively. This hybrid architecture design enables the model to complement both multi-step and single-step reasoning capabilities. Finally, to address the issue of noise in TKGs, we adopt an ensemble learning strategy, combining the predictions of the three modules to reduce the impact of noise on the final prediction results. In the evaluation on six benchmark datasets, our approach generally outperforms existing TKG reasoning models in multi-step and single-step reasoning tasks.

RLGNet: Repeating-Local-Global History Network for Temporal Knowledge Graph Reasoning

TL;DR

RLGNet tackles the challenging problem of extrapolating future events in Temporal Knowledge Graphs by integrating multi-scale historical information through a three-module ensemble: Repeating History, Local History, and Global History. The Local module uses an RNN to capture short-term patterns, the Global module employs an MLP to model long-term trends, and the Repeating module reinforces recurring facts; together they produce a final score via a weighted combination with parameter . The approach is trained with independent losses per module and leverages candidate entity constraints to reduce noise. Empirical results on six benchmark datasets show consistent improvements in both single-step and multi-step reasoning, highlighting the benefit of a multi-scale ensemble design for temporal reasoning tasks.

Abstract

Temporal Knowledge Graph (TKG) reasoning involves predicting future events based on historical information. However, due to the unpredictability of future events, this task is highly challenging. To address this issue, we propose a multi-scale hybrid architecture model based on ensemble learning, called RLGNet (Repeating-Local-Global History Network). Inspired by the application of multi-scale information in other fields, we introduce the concept of multi-scale information into TKG reasoning. Specifically, RLGNet captures and integrates different levels of historical information by combining modules that process information at various scales. The model comprises three modules: the Repeating History Module focuses on identifying repetitive patterns and trends in historical data, the Local History Module captures short-term changes and details, and the Global History Module provides a macro perspective on long-term changes. Additionally, to address the limitations of previous single-architecture models in generalizing across single-step and multi-step reasoning tasks, we adopted architectures based on Recurrent Neural Networks (RNN) and Multi-Layer Perceptrons (MLP) for the Local and Global History Modules, respectively. This hybrid architecture design enables the model to complement both multi-step and single-step reasoning capabilities. Finally, to address the issue of noise in TKGs, we adopt an ensemble learning strategy, combining the predictions of the three modules to reduce the impact of noise on the final prediction results. In the evaluation on six benchmark datasets, our approach generally outperforms existing TKG reasoning models in multi-step and single-step reasoning tasks.
Paper Structure (22 sections, 18 equations, 4 figures, 6 tables)

This paper contains 22 sections, 18 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Model Overview
  5. Numerical Embedding
  6. Local History Module
  7. Global History Module
  8. Repeating History Module
  9. Loss Function
  10. Final Score
  11. Experiments
  12. Setup
  13. Datasets
  14. Evaluation Metrics
  15. Baselines
  16. ...and 7 more sections

Figures (4)

  • Figure 1: The figure illustrates the repeating, local, and global perspectives. The red timestamp indicates that an fact identical to the queried fact occurred at that specific moment in time.
  • Figure 2: The upper half represents the Local History Module, which learns local historical information through KG sequences of adjacent timestamps. The lower half comprises the Repeating and Global History Modules, learning repetitive and global historical information respectively by statistically querying candidate entities. The entities within the dashed box are candidate entities. The left side weights and sums the scores of the three modules to obtain the final prediction score.
  • Figure 3: The impact of $\alpha$ on MRR (in percentage) results.
  • Figure 4: The impact of $\omega$ and $top_k$ on MRR(in percentage) results in ICEWS14 and YAGO.