State Variable Effects in Graphical Event Models

Debarun Bhattacharjya, Dharmashankar Subramanian, Tian GaoIJCAI 2020 · OpenReview

Abstract

Many real-world domains involve co-evolving relationships between events, such as meals and exercise, and time-varying random variables, such as a patient's blood glucose levels. In this paper, we propose a general framework for modeling joint temporal dynamics involving continuous time transitions of discrete state variables and irregular arrivals of events over the timeline. We show how conditional Markov processes (as represented by continuous time Bayesian networks) and multivariate point processes (as represented by graphical event models) are among various processes that are covered by the framework. We introduce and compare two simple and interpretable yet practical joint models within the framework with relevant baselines on simulated and real-world datasets, using a graph search algorithm for learning. The experiments highlight the importance of jointly modeling event arrivals and state variable transitions to better fit joint temporal datasets, and the framework opens up possibilities for models involving even more complex dynamics whenever suitable.

Introduction

In temporal modeling, a fundamental distinction exists between two types of variables. Event labels represent instantaneous, irregular occurrences, such as a user clicking a button or a patient taking medication. In contrast, state variables persist in a specific state and transition between states at irregular intervals, such as a server's status (online/offline) or a patient's blood pressure (low/normal/high).

Historically, machine learning models have treated these dynamics separately. Continuous Time Bayesian Networks (CTBNs) excel at modeling state transitions, while Graphical Event Models (GEMs) are designed for event streams. While recent developments like Event-driven CTBNs (ECTBNs) allow events to trigger state changes, they often fail to capture the reciprocal influence: how state variables affect the arrival rates of events. This research introduces the State Variable Graphical Event Model (SVGEM), a unifying framework that jointly models the bi-directional, continuous-time dynamics between discrete state variables and irregular event arrivals.

Methodology

The SVGEM framework utilizes a directed graph where nodes represent either event labels or state variables. This structure allows for complex, cyclic dependencies over time—for instance, a specific state might increase the likelihood of an event, which in turn triggers a transition to a new state.

Graphical representation of an SVGEM (a) and its expanded Equivalent GEM (b).

A key theoretical contribution is the "Equivalent GEM" trick. The authors demonstrate that any SVGEM can be mathematically converted into a standard GEM by treating every possible state transition as a unique event label. While the SVGEM representation is compact and human-interpretable, this equivalence enables the use of existing, tractable likelihood computation methods for model training.

Within this framework, two practical parametric models were introduced:

  • SVGEM_MP_MP: Event rates depend on the current state of parent variables (Markov) and the recent occurrence of parent events (Proximal).
  • SVGEM_MP_PP: Event rates depend on the recent transitions of parent state variables (Proximal) rather than just their current static state.

The model structure is learned from data using a hill-climbing search algorithm optimized for the Bayesian Information Criterion (BIC) score.

Experimental Evaluation

The models were evaluated across four diverse datasets:

  1. Synthetic: Data generated from a known SVGEM process (5 states, 5 events) to validate learning accuracy.
  2. Tennessee Eastman (TE) Process: Simulated chemical plant data with 16 state variables and 18 event labels representing control actions.
  3. Diabetes: A real-world dataset tracking blood glucose levels (1 state) alongside insulin doses, meals, and exercise (11 events) for 70 patients.
  4. Bitcoin: User ratings on an exchange, modeled with 2 state variables (user goodness/fairness) and 6 event labels.

Key Results and Insights

Experimental results consistently showed that SVGEM variants outperform decoupled baselines, such as independent CTBN and PGEM models.

Log-likelihood comparison showing SVGEM variants outperforming baselines across multiple datasets.

On the Diabetes and TE Process datasets, the SVGEM_MP_PP variant proved particularly effective. This suggests that in biological and physical control systems, a change in state (e.g., a glucose spike) is often a more powerful predictor of future events than the current state alone. Furthermore, the SVGEM framework acts as a true unification, mathematically subsuming standard GEMs, CTBNs, and ECTBNs as special cases.

Conclusion and Future Work

While the SVGEM framework significantly advances joint temporal modeling, it currently relies on discrete state variables, requiring discretization of continuous time-series data. Future research may explore more complex historical dependencies and the integration of more expressive event models, such as Piecewise-Constant Conditional Intensity Models (PCIMs), to further enhance performance in domains driven by intricate event-to-event interactions.

Summary updated

Abstract

Many real-world domains involve co-evolving relationships between events, such as meals and exercise, and time-varying random variables, such as a patient's blood glucose levels. In this paper, we propose a general framework for modeling joint temporal dynamics involving continuous time transitions of discrete state variables and irregular arrivals of events over the timeline. We show how conditional Markov processes (as represented by continuous time Bayesian networks) and multivariate point processes (as represented by graphical event models) are among various processes that are covered by the framework. We introduce and compare two simple and interpretable yet practical joint models within the framework with relevant baselines on simulated and real-world datasets, using a graph search algorithm for learning. The experiments highlight the importance of jointly modeling event arrivals and state variable transitions to better fit joint temporal datasets, and the framework opens up possibilities for models involving even more complex dynamics whenever suitable.
BibTeX
@inproceedings{bhattacharjya2020state,
  author = {Debarun Bhattacharjya and Dharmashankar Subramanian and Tian Gao},
  title = {{State Variable Effects in Graphical Event Models}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=GFSyKkU7lg},
}
Figure