Combining Machine Learning and Queueing Theory for Data-Driven Incarceration-Diversion Program Management

Bingxuan Li, Antonio Castellanos, Pengyi Shi, Amy R. WardAAAI 2024 · DOI

Introduction

Incarceration-diversion programs serve as a vital alternative to traditional prison systems, offering specialized treatments for mental health or substance abuse to break the cycle of recidivism. Despite their potential, managing these programs presents a significant logistical challenge for government agencies. Program managers must balance "sizing" (the number of participants to admit) with "staffing" (the number of case managers required) to ensure community safety and program efficacy. This paper introduces a Decision Support System (DSS) designed to optimize these variables by predicting the program "census"—the total number of active participants and their specific needs over time.

The Management Challenge

Traditional machine learning (ML) models are effective at predicting whether a single individual will succeed or fail in a program. However, they often fall short in operational management because they cannot predict the aggregate system load. Program managers need to know not just who will succeed, but how many people will be in the system at any given time, especially since participants have varying lengths-of-stay (LOS) and outcomes. If a program is undersized, it fails to reach those in need; if it is oversized, it risks "in-program revocations," where overloaded staff cannot prevent violations that lead back to incarceration.

Methodology: A Hybrid Approach

To address these challenges, the researchers developed a framework that integrates ML-based outcome prediction with a queueing-based simulation model. This hybrid approach allows for the capture of temporal dynamics that static ML models miss.

System architecture integrating ML extractors and generators with a queueing-based sampling model.

The framework operates in several stages:

  1. Outcome Prediction: Several models were evaluated, including Logistic Regression and Multi-layer Perceptrons, with Gradient Boosting Trees (GBT) emerging as the most effective for predicting one of four categorical outcomes: Completed, Not Completed, Revoked, or Other.
  2. Queueing Simulation: The system simulates individual flow through the program based on historical arrival rates. For each arrival, the model samples a predicted outcome and a corresponding LOS.
  3. Handling Right-Censoring: A significant technical hurdle in criminal justice data is "right-censoring," where participants are still in the program at the time of data collection. The authors implemented a two-stage sampling approach: first predicting the likelihood of being right-censored, and then sampling LOS from empirical distributions based on county, year, race, and outcome.

Data and Implementation

The study utilized longitudinal, de-identified data from four Illinois counties (Cook, DuPage, Peoria, and Will) spanning 2011 to 2022. The dataset included demographics, risk assessment scores, and criminal history for hundreds of participants per county. The resulting DSS was deployed as a web application, enabling managers to run counterfactual "what-if" scenarios regarding admission rates and staffing levels.

Results and Operational Impact

The GBT model achieved a One-vs-Rest (OVR) ROC AUC of 0.718. More importantly, the simulation's ability to predict the total participant census was highly accurate. The Mean Absolute Percentage Error (MAPE) for predicting the total number of participants remained between 9.40% and 21.21% even when using predicted outcomes.

Simulated census tracking actual participant numbers over time, demonstrating the model's temporal accuracy.

Operational insights from the DSS revealed critical nuances in program management. For example, the system identified that while Hispanic participants often have higher completion rates, they also tend to have longer LOS. This insight signals a specific need for more Spanish-speaking case managers to handle the higher census within that subgroup. Furthermore, counterfactual analysis suggested that the relationship between admission rates and societal cost follows a U-shaped curve, helping managers find the "sweet spot" for program eligibility.

Conclusion and Future Work

By combining the predictive power of ML with the structural modeling of queueing theory, this framework provides a robust tool for incarceration-diversion management. While the model currently assumes future arrivals will mirror historical patterns, future work aims to incorporate arrival forecasting and broader socio-determinants to further refine predictions. This data-driven approach offers a path toward more efficient, safe, and impactful diversion programs.

Summary updated

BibTeX
@inproceedings{li2024combining,
  author = {Bingxuan Li and Antonio Castellanos and Pengyi Shi and Amy R. Ward},
  title = {{Combining Machine Learning and Queueing Theory for Data-Driven Incarceration-Diversion Program Management}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30330},
}
Figure