Order-Dependent Event Models for Agent Interactions

Debarun Bhattacharjya, Tian Gao, Dharmashankar SubramanianIJCAI 2020 · OpenReview

Abstract

In multivariate event data, the instantaneous rate of an event's occurrence may be sensitive to the temporal sequence in which other influencing events have occurred in the history. For example, an agent’s actions are typically driven by preceding actions taken by the agent as well as those of other relevant agents in some order. We introduce a novel statistical/causal model for capturing such an order-sensitive historical dependence, where an event’s arrival rate is determined by the order in which its underlying causal events have occurred in the recent past. We propose an algorithm to discover these causal events and learn the most influential orders using time-stamped event occurrence data. We show that the proposed model fits various event datasets involving single as well as multiple agents better than baseline models. We also illustrate potentially useful insights from our proposed model for an analyst during the discovery process through analysis on a real-world political event dataset.

Introduction

In fields ranging from healthcare to high-frequency trading and geopolitics, researchers are increasingly tasked with analyzing "event streams"—sequences of time-stamped actions. To model these, Graphical Event Models (GEMs) are often employed. These models use conditional intensity functions to calculate the instantaneous rate of an event occurring based on its history.

However, most existing GEMs are "order-neutral." For example, a Proximal GEM (PGEM) might observe a 5-day window and ask: "Did a 'Negative Statement' and a 'Military Threat' occur?" If both occurred, the model raises the probability of a 'Declaration of War.' In reality, the order of these events is critical: a threat followed by a statement might signal de-escalation, whereas a statement followed by a threat signals escalation. This paper introduces the Ordinal Graphical Event Model (OGEM) to address this inability to distinguish between distinct causal histories.

The Ordinal Graphical Event Model (OGEM)

The core innovation of the OGEM is its ability to capture the nuanced causal impact of event sequences. Unlike standard models that only check for the presence of parent events, the OGEM looks at the specific permutation (the order) in which they occurred within a defined time window ($w$).

Figure 2: Illustration of event sequences and the corresponding graphical model structure where rates depend on the order of parent events.

To manage the complexity of real-world data where events may repeat, the authors utilize masking functions ($\phi$). These functions retain either only the first or the last occurrence of a specific event type to determine the sequence. The model's structure is learned via an efficient greedy forward-backward score-based search using the Bayesian Information Criterion (BIC), followed by Maximum Likelihood Estimation (MLE) to calculate intensity rates for every possible order of parent events.

Experimental Evaluation

The OGEM was evaluated against four diverse real-world datasets: socio-political events (ICEWS), ICU electronic health records (Mimic-II), diabetes management data, and professional career paths (LinkedIn).

Performance Comparison

In 6 out of 7 experimental setups, OGEMs achieved a higher log-likelihood on test data than the order-neutral PGEMs. This confirms that order-dependence is a measurable and significant feature of these complex datasets. Furthermore, on the ICEWS political data, OGEMs performed comparably to—and sometimes better than—state-of-the-art Neural Hawkes Processes (NHP), while maintaining much higher interpretability.

Table 4: Log-likelihood comparison across datasets. OGEM consistently outperforms the order-neutral PGEM and remains competitive with neural models.

Causal Discovery

Beyond predictive accuracy, the OGEM provides actionable insights. In the ICEWS dataset, the model identified specific "escalation" sequences. For instance, in Venezuela, a "Citizen Statement" followed by a "Government Statement" resulted in a significantly higher rate of "Police-Citizen Fighting" than the reverse order. This level of transparency allows analysts to rank event sequences by their influential power, a feature missing in "black box" neural models.

Conclusion

The Ordinal Graphical Event Model represents a significant step forward in event stream analysis by proving that "order matters." While the model faces challenges regarding "state-space explosion" as the number of parent events increases, its superior fit and interpretability make it a powerful tool for causal discovery. Future work will focus on automating hyperparameter learning for time windows and developing more compact representations to handle even more complex agent interactions.

Summary updated

Abstract

In multivariate event data, the instantaneous rate of an event's occurrence may be sensitive to the temporal sequence in which other influencing events have occurred in the history. For example, an agent’s actions are typically driven by preceding actions taken by the agent as well as those of other relevant agents in some order. We introduce a novel statistical/causal model for capturing such an order-sensitive historical dependence, where an event’s arrival rate is determined by the order in which its underlying causal events have occurred in the recent past. We propose an algorithm to discover these causal events and learn the most influential orders using time-stamped event occurrence data. We show that the proposed model fits various event datasets involving single as well as multiple agents better than baseline models. We also illustrate potentially useful insights from our proposed model for an analyst during the discovery process through analysis on a real-world political event dataset.
BibTeX
@inproceedings{bhattacharjya2020order,
  author = {Debarun Bhattacharjya and Tian Gao and Dharmashankar Subramanian},
  title = {{Order-Dependent Event Models for Agent Interactions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=5V8lnsTHQfs},
}
Figure