Learning the Influence between Partially Observable Processes using Scorebased Structure Learning

Ritesh Ajoodha, Benjamin RosmanAdvances 2020 · DOI

Abstract: The difficulty of learning the underlying structure between processes is a common task found throughout the sciences, however not much work is dedicated towards this problem. In this paper, we attempt to use the language of structure learning to address learning the dynamic influence network between partially observable processes represented as dynamic Bayesian networks. The significance of learning an influence network is to promote knowledge discovery and improve on density estimation in the temporal space. We learn the influence network, defined by this paper, by learning the optimal structure for each process first, and thereafter apply redefined structure learning algorithms for temporal models. Our procedure builds on the language of probabilistic graphical model representation and learning. This paper provides the following contributions: we (a) provide a definition of influence between stochastic processes represented by dynamic Bayesian networks; (b) expand on the conventional structure learning literature by providing a structure score and learning procedure for temporal models; and (c) introduce the notion of a structural assemble which is used to associate two stochastic processes represented by dynamic Bayesian networks.

Introduction

In various scientific domains, researchers encounter stochastic processes that evolve over time. A significant challenge lies in identifying the dynamic influence networks (DINs) that govern how these processes interact. Traditionally, these systems are either modeled in isolation—ignoring cross-process dependencies—or as a single, monolithic "black box," which obscures the internal logic of individual processes. Furthermore, many real-world systems are only partially observable, adding another layer of complexity to structure learning, which is known to be an NP-hard problem.

Methodology

The authors propose a score-based structure learning framework to address these challenges. The methodology is built upon four key pillars:

  1. Dynamic Influence Networks (DINs): A formal mathematical definition for the relationships between multiple Dynamic Bayesian Networks (DBNs).
  2. The d-BIC Score: A "dynamic" adaptation of the Bayesian Information Criterion. This scoring metric balances model fit with complexity, penalizing overly dense networks to prevent the identification of spurious connections.
  3. Structural Assemble: A template-based approach to defining associations between temporal models.
  4. Greedy Structure Search (GESS): An iterative algorithm that navigates the vast search space of potential network structures by making local improvements to the d-BIC score.

To handle hidden variables in partially observable systems, the framework integrates Expectation-Maximization (EM), which estimates missing data while simultaneously refining the network structure.

The proposed learning procedure for Dynamic Influence Networks, illustrating the iterative GESS and EM workflow.

Experimental Evaluation

The framework was evaluated using synthetic datasets generated from a known "ground-truth" DIN. This setup allowed for precise measurement of the algorithm's ability to recover the true underlying structure. The experimental parameters included 10 DBNs, each containing 5 observable and 3 latent variables, tracked over 5 time-slices.

The performance was assessed by measuring the Relative Entropy between the learned model and the ground truth across varying sample sizes (up to 1,400 instances).

Key Results and Discussion

The experimental results demonstrate the effectiveness of the d-BIC and d-AIC scoring metrics compared to baseline approaches.

Experimental parameters and performance results showing Relative Entropy vs. Number of Samples for various structure learning methods.

Key findings include:

  • Superior Accuracy: Penalty-based scores (d-BIC and d-AIC) significantly outperformed random and tree-based structures. As the sample size increased, the performance of the d-BIC-learned structure converged toward that of the "True Structure."
  • The "No Structure" Threshold: For very small datasets (fewer than 250 samples), assuming no influence between processes often yielded more accurate inferences than attempting to learn a complex structure from insufficient data.
  • Complexity Management: The d-BIC score successfully maintained network sparsity, ensuring the learned models remained computationally tractable for future predictions.

Conclusion and Future Work

This research provides a robust foundation for learning influences between partially observable stochastic processes. While the framework shows high accuracy, the authors note that "I-equivalence" means multiple structures may fit the data equally well, even if they aren't the physical ground truth. Future work will focus on scaling the algorithm for even larger systems and validating the approach on real-world datasets, such as IoT sensor networks or educational tracking systems.

Summary updated

Abstract

The difficulty of learning the underlying structure between processes is a common task found throughout the sciences, however not much work is dedicated towards this problem. In this paper, we attempt to use the language of structure learning to address learning the dynamic influence network between partially observable processes represented as dynamic Bayesian networks. The significance of learning an influence network is to promote knowledge discovery and improve on density estimation in the temporal space. We learn the influence network, defined by this paper, by learning the optimal structure for each process first, and thereafter apply redefined structure learning algorithms for temporal models. Our procedure builds on the language of probabilistic graphical model representation and learning. This paper provides the following contributions: we (a) provide a definition of influence between stochastic processes represented by dynamic Bayesian networks; (b) expand on the conventional structure learning literature by providing a structure score and learning procedure for temporal models; and (c) introduce the notion of a structural assemble which is used to associate two stochastic processes represented by dynamic Bayesian networks.
BibTeX
@inproceedings{ajoodha2020learning,
  author = {Ritesh Ajoodha and Benjamin Rosman},
  title = {{Learning the Influence between Partially Observable Processes using Scorebased Structure Learning}},
  booktitle = {Advances},
  year = {2020},
  doi = {10.25046/aj050503},
}
Figure