Cause-Effect Association between Event Pairs in Event Datasets

Debarun Bhattacharjya, Tian Gao, Nicholas Mattei, Dharmashankar SubramanianIJCAI 2020 · OpenReview

Abstract

Causal discovery from observational data has been intensely studied across fields of study. In this paper, we consider datasets involving irregular occurrences of various types of events over the timeline. We propose a suite of scores and related algorithms for estimating the cause-effect association between pairs of events from such large event datasets. In particular, we introduce a general framework and the use of conditional intensity rates to characterize pairwise associations between events. Discovering such potential causal relationships is critical in several domains, including health, politics and financial analysis. We conduct an experimental investigation with synthetic data and two real-world event datasets, where we evaluate and compare our proposed scores using assessments from human raters as ground truth. For a political event dataset involving interaction between actors, we show how performance could be enhanced by enforcing additional knowledge pertaining to actor identities.

Introduction

Many real-world domains, from healthcare to international relations, generate event datasets where occurrences are irregular and "spiky." Unlike traditional time series or i.i.d. data, these event streams require specialized methods to uncover causal relationships. Existing approaches, often adapted from Natural Language Processing (NLP) co-occurrence metrics, frequently rely on ad-hoc parameters and fail to account for the continuous-time nature of event rates. This paper introduces a principled framework based on Point Processes and Conditional Intensity Ratios (CIR) to address these challenges.

Methodology

The core of our approach is the comparison of event "intensities"—the rate at which an event occurs—under different conditions. We propose three primary scores:

  1. $CIR_B$ (Baseline): Compares the intensity of an effect $x$ following a potential cause $y$ to the global average intensity of $x$.
  2. $CIR_C$ (Contrastive): Compares the intensity of $x$ following $y$ to the intensity when $y$ has not recently occurred.
  3. $CIR_M$ (Multivariate): Uses structure learning to identify "parent" events for $x$, calculating the causal score for a pair while controlling for other influences. This multivariate approach is designed to filter out spurious correlations.

Experimental Evaluation

We evaluated our proposed scores against adapted NLP baselines ($ECD_E$ and $NST_E$) using synthetic and real-world datasets.

Synthetic Data Results

On synthetic traces generated from known Directed Acyclic Graphs (DAGs), the CIR-based models consistently outperformed the baselines. As shown in Figure 4, the intensity-based scores maintain superior performance across various window sizes ($w=7, 15, 30$), demonstrating the robustness of the continuous-time framework.

Performance comparison on synthetic data across different window sizes (7, 15, 30). CIR-based models consistently outperform NLP-adapted baselines in identifying true causal pairs.

Real-World Applications

In a medical dataset tracking diabetes management (glucose levels, insulin doses, and exercise), we measured performance using Hits@K, which counts how many true causal links appear in the top $K$ results. The $CIR_M$ model, by accounting for multivariate influences, significantly outperformed other methods.

Hits@K results for the Diabetes dataset. CIRM achieves the highest number of correctly identified causal links among the top K results.

For the ICEWS political dataset, we found that incorporating domain-specific constraints—such as requiring cause and effect events to share a common actor—further improved the alignment between algorithmic scores and human plausibility ratings.

Discussion and Conclusion

Our results highlight the importance of treating time as continuous and accounting for multivariate parent sets in causal discovery for event streams. While the current framework uses fixed influence windows, future work will explore automated window parameter learning and the inclusion of inhibitory causal effects. This research provides a scalable, mathematically grounded foundation for analyzing complex event-driven systems.

Summary updated

Abstract

Causal discovery from observational data has been intensely studied across fields of study. In this paper, we consider datasets involving irregular occurrences of various types of events over the timeline. We propose a suite of scores and related algorithms for estimating the cause-effect association between pairs of events from such large event datasets. In particular, we introduce a general framework and the use of conditional intensity rates to characterize pairwise associations between events. Discovering such potential causal relationships is critical in several domains, including health, politics and financial analysis. We conduct an experimental investigation with synthetic data and two real-world event datasets, where we evaluate and compare our proposed scores using assessments from human raters as ground truth. For a political event dataset involving interaction between actors, we show how performance could be enhanced by enforcing additional knowledge pertaining to actor identities.
BibTeX
@inproceedings{bhattacharjya2020cause,
  author = {Debarun Bhattacharjya and Tian Gao and Nicholas Mattei and Dharmashankar Subramanian},
  title = {{Cause-Effect Association between Event Pairs in Event Datasets}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=vTUAofnKPvY},
}
Figure