# Capturing Event Promoting and Inhibiting Associations by Learning Semantic Chains

# Abstract

Abstract—Analyzing the associations among events is crucial for understanding the interaction mechanisms within event data. In real-life applications, people are particularly interested in promoting and inhibiting associations between events, which are influenced by contextual information and hidden causality association. However, existing approaches to event association analysis often focus on causal, temporal, and similarity relationships, addressing only one{some} of these factors. No single approach has successfully integrated all these aspects. In this paper, we focus on analyzing promoting and inhibiting associations between events. We quantify promoting and inhibiting associations based on the amount of contextual information and hidden causal associations between events derived from past observations. We propose a novel framework, EPIR, for promoting and inhibiting analysis. Specifically, we first propose an event contextual learning module, MCAR, to capture the temporal, spatial, and semantic associations between events. Then, we construct a causal schema-based model, MCSR, to analyze the associations between events. MCSR treats the causal schema as a latent variable and simultaneously trains a schema generator and a reasoning predictor. Extensive tests on event datasets are conducted to demonstrate the high effectiveness of our approach.

# Introduction

Event association analysis aims to identify specific associations between events within data [1], which is crucial for data analysis-based applications, including question answering [2] and timeline construction [3]. In practical applications, researchers are particularly interested in analyzing the positive and negative effects that the development of one event has on other events [4], [5]. Currently, event association analysis approaches focus on temporal relations [6], sub-event relations [7], causal relationships [8], [9], and other aspects.{None of these approaches can directly describe these interactions}, particularly the negative influences between events. Effectively analyzing the promoting and inhibiting associations greatly aids various applications, including emergency management, risk identification, and others.

Existing event association analysis approaches focus on event temporal relations, sub-event relations, and causal relationships.

事件时序关系、子事件关系和因果关系。

paramount : adj. 至为重要的，首要的；至高无上的，权力最大的

Intuitively, promoting and inhibiting associations highlight the interactive effects between events and{along with} their attributes, while other types of relationships focus on the attributes of the events or the structured connections between events. More specifically, such interactive associations include two factors. (1) Contextual information. =={一种潜在的关系信息，也可称为语境}==An event may be correlated not only with contemporaneous events but also with past events. Existing approaches to event association analysis often focus on spatial, temporal, and similarity relationships, addressing only some of these factors. It is challenging to integrate all these aspects into contextual learning. (2) Hidden causal associations, where the association between events is directly reflected in the causal relationships among their attributes. Such interaction are revealed through an implicit causal chain. Although statistical learning can infer these associations, existing approaches may not fully capture hidden causal relationships. Consequently, there is a pressing need for an accurate algorithm to analyze promoting and inhibiting relationships.

implicit causal link 隐藏因果链

# Challenge

Challenges. Analyzing promoting and inhibiting associations is challenging {a challenging task / 或其他}. Existing approaches have focused on event association from textual data or graphs, but they cannot express {改成清晰表达可能好点} promoting and inhibiting associations in structured data. Therefore, several open questions remain. How can we define promoting and inhibiting associations between events? More specifically, how can we define contextual associations and hidden causal associations between events? How can we quantify promoting and inhibiting associations using con- textual information and hidden causal associations? Moreover, can we extract the promoting and inhibiting associations between events? That is, can we extract contextual information and hidden causal associations between events?

# Novelty

Novelty. As opposed to prior work, our approach introduces the following innovations. (1) We introduce promoting and inhibiting relationships and quantify them by incorporating event contextual associations and hidden causal associations. (2) Unlike graph-based methods that focus on the interactions between event attributes and structures, we apply an event hidden causal schema to extract causal associations within event chains. This enables us to identify hidden causal schema. (3) Unlike textualbased methods that focus on the co-occurrence of event phrase pairs, our approach applies event contextual associations using an contextual learning module. This enables us to analyze the contextual relationships between event attributes based on their semantic correlation rather than their frequency.

# implicit reasoning rules

generate

规则概率，

概率分布，grounding 确定 path 的得分，rule 的得分

选 10 个然后然后计算得分

score

计算 p = softmax (score) 则对查询事件（h,*,t）,generator 会根据事件库中不同的 label (包括 t) 生成一个分布并依据这个分布计算 score，从而得到不同 label 的得分如果 t 的得分是最大的则认为其具有隐藏因果链关系，反之则认为不具有隐藏因果链关系。

根据搜索结果，以下是 2023 年和 2024 年数据库或机器学习领域中关于因果、时序、子事件的相关论文总结：

1. 论文标题: Causal Discovery from Subsampled Time Series with Proxy Variable

   • 解决的问题：在时间序列数据中，由于采样时间分辨率不足，导致因果结构识别困难。该论文提出了一种基于代理变量的时序因果发现理论，以解决非参意义下的因果图识别性保证问题。
   • 解决方法：论文设计了一种基于条件独立性测试的因果发现算法，并在仿真实验和阿尔茨海默疾病分析中取得了优秀表现。
   • 论文链接: arXiv

2. 论文标题: Periodicity Decoupling Framework for Long-term Series Forecasting

   • 解决的问题：长期序列预测中的时间变化建模问题。
   • 解决方法：提出了一种新颖的周期性解耦框架（PDF），通过捕获 2D 时间变化来进行长期序列预测。
   • 论文链接: OpenReview

3. 论文标题: Self-Supervised Contrastive Forecasting

   • 解决的问题：长期序列预测问题。
   • 解决方法：提出了一种自监督对比学习方法，通过捕获 2D 时间变化建模来进行长期序列预测。
   • 论文链接: OpenReview

4. 论文标题: TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series

   • 解决的问题：多元时间序列预测问题。
   • 解决方法：提出了一种文本原型对齐嵌入方法，以激活大型语言模型（LLM）在时间序列预测中的能力。
   • 论文链接: OpenReview

5. 论文标题: CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns

   • 解决的问题：时间序列预测中的周期性模式建模问题。
   • 解决方法：提出了 CycleNet，通过对周期模式进行建模增强时间序列预测。
   • 论文链接: arXiv

6. Causal Discovery from Subsampled Time Series with Proxy Variable

   • 解决了什么问题：该论文提出了一种基于代理变量（Proxy Variables）的时序因果发现理论，解决了非参意义下的因果图识别性保证问题。特别是在数据采样时间分辨率不足的情况下，如何准确识别因果结构。
   • 怎么解决的：论文设计了一种基于条件独立性测试的因果发现算法，并在仿真实验和阿尔茨海默疾病分析中取得了优秀表现。
   • 论文链接：arXiv

7. CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns

   • 解决了什么问题：该论文旨在通过建模周期模式来增强时间序列预测，特别是在长时预测和周期建模方面。
   • 怎么解决的：提出了 CycleNet 模型，专注于捕捉时间序列中的周期性模式，以提高预测的准确性。
   • 论文链接：arXiv

8. From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection

   • 解决了什么问题：该论文探讨了如何将事件分析整合到基于大型语言模型（LLM）的时间序列预测中，以提高预测的准确性和鲁棒性。
   • 怎么解决的：提出了一种新的方法，通过分析新闻事件来增强时间序列预测，使得模型能够更好地理解和预测未来趋势。
   • 论文链接：arXiv

9. Learning diverse causally emergent representations from time series data

   • 解决了什么问题：该论文关注于从时间序列数据中学习多样化的因果涌现表示。
   • 怎么解决的：提出了一种方法来识别和学习时间序列数据中的因果关系，这对于理解和预测复杂系统的行为至关重要。
   • 论文链接：

10. Robust agents learn causal world models

    • 解决了什么问题：该论文探讨了因果推理在鲁棒和通用智能中的作用，并回答了智能体是否必须学习因果模型才能推广到新领域的问题。论文指出，任何能够在大量分布偏移下满足遗憾界的智能体必须已经学习了数据生成过程的近似因果模型，对于最优智能体而言，这一模型会收敛到真实的因果模型。
    • 怎么解决的：论文通过理论分析，表明了学习因果模型对于智能体在不同分布的数据上表现良好是必要的。论文讨论了这一结果对于迁移学习和因果推断等研究领域的影响，从而强调了因果模型在智能体鲁棒性中的重要性。

11. Causally Aligned Curriculum Learning

    • 解决了什么问题：该论文研究了在强化学习（RL）中，由于环境包含未观测到的混杂因素，导致传统的课程学习假设（即最优决策规则在源任务和目标任务之间保持不变）不成立的问题。论文通过因果视角来研究课程 RL 问题，并提出了一个因果对齐的课程学习方法。
    • 怎么解决的：论文首先推导出一个充分的图形条件，用以表征因果对齐的源任务，即在这些任务中最优决策规则的不变性成立。接着，论文开发了一个高效的算法来生成一个因果对齐的课程，前提是提供了目标环境的定性因果知识。最后，论文通过在混杂环境中的实验验证了所提出方法的有效性。

Robust agents learn causal world models

Causally Aligned Curriculum Learning