💻 computer science

Classification of High-Risk Paratransit Drivers Using Ensemble Machine Learning

This study utilizes a suite of machine learning algorithms, with Logistic Regression emerging as the most effective model, to classify high-risk paratransit drivers in Gazipur, Bangladesh, based on socioeconomic and operational factors, thereby offering transportation authorities a data-driven tool for targeted safety interventions.

Original authors: Md Sifat Bin Siraj, Shumaila Noor, Md Emon Miah, Md Jubayadul Islam, Saifullah Mahmud, Jannatul Ferdous, Sk. Md. Ahaduz

Published 2026-08-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Md Sifat Bin Siraj, Shumaila Noor, Md Emon Miah, Md Jubayadul Islam, Saifullah Mahmud, Jannatul Ferdous, Sk. Md. Ahaduz

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a city as a giant, living organism. In many developing cities, the "formal" transport system—like big buses on fixed schedules—is often the skeleton, but it's missing a few bones. When the skeleton can't reach every corner, a flexible, informal layer of small vehicles (rickshaws, motorized three-wheelers, and easy bikes) steps in to fill the gaps. This informal layer is the city's immune system and its emergency response team rolled into one; it keeps people moving even when floods hit, strikes happen, or fuel runs out. But just like any immune system, it has weak spots. The biggest variable in this system isn't the road or the weather; it's the driver. For decades, safety experts have tried to figure out who is likely to crash using old-school math that counts accidents like tally marks on a wall. But counting past accidents is like trying to predict a storm by looking at puddles after the rain has already fallen. The real challenge is spotting the "storm clouds" before the first drop hits. This is where a new branch of science called machine learning steps in. Think of machine learning not as a crystal ball, but as a super-smart detective that looks at a driver's entire life story—their income, their vehicle, their habits—to guess who might be in trouble tomorrow, rather than just punishing them for what happened yesterday.

This paper is about that detective work. A team of researchers went to Gazipur, Bangladesh, a bustling industrial city where thousands of these informal drivers work, to see if they could build a better "risk radar." They interviewed 507 drivers, asking about their daily lives, their earnings, their vehicles, and their crash history. Instead of just counting how many times someone crashed, they used five different machine learning algorithms (digital brain patterns) to sort these drivers into two groups: "Safe" and "High-Risk." They wanted to see if these digital detectives could find the dangerous drivers better than the old math methods could.

The results were a bit surprising and very practical. The researchers found that the simplest tool in their kit, a model called Logistic Regression, actually did the best job. It managed to correctly identify about 57% of the truly risky drivers, which was a huge improvement—about 40% better at catching them—compared to the traditional counting methods. The old methods were so focused on being precise that they missed a lot of the actual troublemakers. The new model, however, was better at spotting the potential hazards, even if it occasionally flagged a safe driver by mistake (which the researchers decided was a safer trade-off than missing a dangerous one).

The most interesting part of the story is what the model found out about why drivers are risky. The data shattered a common assumption: that the more experience a driver has, the safer they are. In this study, the "High-Risk" group actually had more driving experience (averaging 8.6 years) than the "Safe" group (averaging 6.2 years). The researchers suggest this might be because experienced drivers get overconfident or complacent, thinking they know the roads too well to make mistakes. It's like a veteran gamer who stops checking their corners because they think they've seen it all before.

Another massive red flag was the type of vehicle. Drivers operating motorized three-wheelers (like CNG auto-rickshaws) were roughly three times more likely to be high-risk than those driving non-motorized rickshaws or easy bikes. The machine learning model treated the vehicle type as the single most important clue, more important than age or education. It turns out that the speed and power of the motorized vehicles, combined with the pressure to earn more money, create a perfect storm for accidents.

So, what does this mean for the real world? The paper suggests that city officials shouldn't just wait for accidents to happen and then punish the drivers. Instead, they can use this "risk radar" to check drivers before they renew their licenses. If a driver has a motorized vehicle, high daily income (which might mean they are rushing), and lots of experience, the system can flag them for a refresher course or a safety check. The researchers argue that this approach turns a boring administrative list of license holders into a living safety map. It doesn't require expensive cameras or sensors; it just uses the information the city already has. By focusing on the specific mix of vehicle type, experience, and economic pressure, cities can protect their most vulnerable transport layer, keeping the "immune system" of the city strong and ready to handle whatever comes next.

Technical Summary: Classification of High-Risk Paratransit Drivers Using Ensemble Machine Learning

Problem Statement
In developing urban centers like Gazipur, Bangladesh, paratransit systems (e.g., auto-rickshaws, easy bikes) serve as a critical buffer for urban transport resilience, absorbing demand that formal networks cannot meet. However, the safety of this informal fleet is poorly characterized. Conventional regulatory systems struggle to address safety issues because driver-level risk is often assessed using traditional statistical methods (e.g., Poisson or Negative Binomial regression) that model accident frequency rather than predicting individual risk categories. These linear models fail to capture non-linear interactions between driver behavior, socioeconomic factors, and vehicle types, limiting the ability of authorities to proactively identify high-risk drivers for targeted intervention. Furthermore, crash data in this sector is inherently imbalanced, with safe drivers vastly outnumbering those with crash histories, complicating the training of predictive models.

Methodology
The study utilized a dataset of 507 paratransit drivers in Gazipur, collected via structured face-to-face interviews between July and December 2021. Drivers were categorized into two groups based on self-reported crash history: "Safe" (one accident, $n=386$ ) and "Risky" (two or more accidents, $n=121$ ).

The research employed a comparative machine learning approach involving five classification algorithms:

Logistic Regression: A baseline interpretable model with L2 regularization.
Random Forest: An ensemble of 200 decision trees using bootstrap aggregating.
XGBoost: A sequential gradient boosting algorithm with regularization.
Gradient Boosting: A sequential tree-fitting approach.
Stacking Ensemble: A meta-learner combining Logistic Regression, Random Forest, and XGBoost.

To address the class imbalance (76% safe vs. 24% risky), the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set to generate synthetic risky examples, balancing the training data to a 270:270 ratio. The dataset was split into 70% training and 30% testing using stratified sampling.

The study compared these machine learning classifiers against a traditional Poisson regression model (converted to binary classification for comparison). Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, AUC-ROC, and the Matthews Correlation Coefficient (MCC), with F1-Score designated as the primary metric due to the importance of balancing false positives and false negatives in safety contexts. Feature importance was analyzed using Gini importance, gain importance, and standardized regression coefficients.

Key Results

Model Performance: Logistic Regression emerged as the best-performing model, achieving an F1-score of 0.483 and a Recall of 0.568. This outperformed all ensemble methods (Random Forest F1=0.406, Stacking F1=0.394, XGBoost F1=0.369).
Comparison with Traditional Methods: The Logistic Regression model demonstrated a 12.6% improvement in F1-score and a 40% improvement in Recall compared to the traditional Poisson regression model. While Poisson regression showed slightly higher precision, its significantly lower recall (0.405) meant it failed to identify a substantial portion of high-risk drivers, a critical failure for proactive safety management.
Feature Importance:
- Vehicle Type: The operation of motorized three-wheelers (CNG/leguna/tempo) was identified as the single most dominant predictor of high-risk status, associated with roughly three times the crash risk of non-motorized modes.
- Experience: Contrary to conventional safety assumptions, driving experience was positively correlated with risk. Risky drivers had significantly more years of experience (8.6 years) than safe drivers (6.2 years), suggesting that accumulated experience may lead to complacency or overconfidence rather than enhanced safety.
- Socioeconomic Factors: Higher daily income and driving speed were also significant predictors of risk.
Calibration: While the Stacking Ensemble showed the best calibration (lowest Brier score), Logistic Regression provided a superior balance of discrimination and interpretability.

Significance and Contributions
The paper claims three primary contributions to transportation safety literature:

Methodological Advancement: It validates that machine learning classification models, particularly interpretable ones like Logistic Regression, outperform traditional count regression models for the specific policy goal of identifying high-risk individuals. The study demonstrates that classification frameworks are better suited for binary policy decisions (e.g., "retrain this driver") than frequency-based models.
Handling Data Imbalance: The study successfully applies SMOTE to driver safety data, enabling robust model training despite the scarcity of crash records in the informal sector.
Policy-Relevant Insights: The research challenges the assumption that experience equates to safety, revealing that experienced drivers in the paratransit sector may pose higher risks. It further highlights that vehicle type is a more significant risk factor than individual driver demographics.

Significance for Governance
The authors argue that this classification framework provides transportation authorities with an actionable tool for targeted safety interventions. By utilizing existing administrative data (vehicle type, income, experience), authorities can generate risk scores to prioritize resources toward high-risk drivers without requiring expensive telematics or new infrastructure. The study suggests that integrating risk screening into the permit renewal cycle could transform routine administrative data into a standing indicator of urban transport resilience, supporting Sustainable Development Goals (SDGs) 9 and 11.2 by improving the dependability of the informal transport layer that cities rely on during disruptions.

Technical Summary: Classification of High-Risk Paratransit Drivers Using Ensemble Machine Learning

More like this