Online Modeling and Monitoring of Dependent Processes under Resource Constraints

Tanapol Kosolwattana, Huazheng Wang, Ying Lin

Abstract

Monitoring a population of dependent processes under limited resources is critical for abnormal events detection. A novel online collaborative learning method is proposed to adaptively allocate the resources for exploitation of high-risk processes and exploration of dependent dynamics. Efficiency of the proposed method is proved through theoretical analysis and experiments.

Introduction

In many real-world scenarios, such as healthcare systems or industrial sensor networks, we face the challenge of monitoring a large population of dynamic processes with limited resources. For instance, a clinic might have 1,000 patients at risk for Alzheimer’s disease but only enough capacity to evaluate 100 of them per month. To prevent severe deterioration, it is vital to prioritize the highest-risk individuals. However, because disease progression is dynamic and often unknown initially, we encounter a classic "exploration-exploitation" dilemma.

Exploitation involves monitoring patients known to be in poor health, while exploration focuses on patients with uncertain conditions to improve predictive models. Traditional methods often treat individuals as independent or ignore predictive uncertainty. This research introduces a framework that leverages dependent processes—the principle that individuals with similar profiles (e.g., genetics or demographics) share similar trajectories—to optimize monitoring decisions under strict constraints.

The Online Collaborative Learning Framework

The authors propose the Online Collaborative Learning (OCL) framework, which integrates collaborative modeling with a multi-armed bandit strategy.

The proposed online collaborative learning framework for health monitoring.

The framework consists of two primary components:

  1. Collaborative Health Modeling: Instead of isolated models, the system assumes a few "canonical" progression patterns. Each individual's trajectory is modeled as a weighted combination of these patterns, effectively sharing information across the population via matrix factorization.
  2. CL-UCB Algorithm: The "Collaborative Learning-based Upper Confidence Bound" strategy calculates a priority score for each unit. This score balances the predicted health state (exploitation) with model uncertainty (exploration). By selecting the $M$ units with the highest scores, the system optimizes both immediate care and long-term model accuracy.

To maintain real-time performance, an Alternating Least Squares (ALS) approach updates model parameters as new data arrives, allowing the system to adapt to changing population dynamics.

Experimental Evaluation

The framework was evaluated using both synthetic simulations and real-world data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The ADNI dataset included longitudinal Mini-Mental State Examination (MMSE) scores, using baseline risk factors like ApoE genotypes and brain volume to establish initial similarities.

Key Results

  • Reduced Regret: In both synthetic and real-world tests, the CL-UCB algorithm consistently demonstrated lower cumulative regret—the cost of sub-optimal monitoring choices—compared to benchmarks like LinUCB and GOB.Lin.
  • Faster Convergence: Parameter estimation accuracy improved significantly faster with CL-UCB, meaning the system identified health trajectories earlier.
  • Scalability: Sensitivity analysis revealed that while traditional methods' errors grew linearly with population size, the collaborative approach remained robust, making it suitable for large-scale applications.

Discussion and Future Work

The core insight of this research is that accounting for process dependency acts as a powerful regularizer. By assuming population dynamics can be represented by a small number of canonical models ($K \ll N$), the algorithm reduces individual-level noise and requires less data for high accuracy. Furthermore, the model can incorporate prior knowledge, such as medical history, via a graph Laplacian matrix.

While the current model focuses on linear progression, future work aims to extend the framework to nonlinear dynamics using kernel regressions. Additionally, the authors plan to explore dynamic resource allocation, where the total monitoring budget can fluctuate based on the population's overall risk level.

Summary updated

Abstract

Monitoring a population of dependent processes under limited resources is critical for abnormal events detection. A novel online collaborative learning method is proposed to adaptively allocate the resources for exploitation of high-risk processes and exploration of dependent dynamics. Efficiency of the proposed method is proved through theoretical analysis and experiments.
BibTeX
@misc{kosolwattana2023online,
  author = {Tanapol Kosolwattana and Huazheng Wang and Ying Lin},
  title = {{Online Modeling and Monitoring of Dependent Processes under Resource Constraints}},
  year = {2023},
  eprint = {2307.14208},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.14208},
}

Loading related content...

Figure