A newly published investigation into road safety governance indicates that previous risky driving habits and past traffic offenses are the predominant factors predicting whether motorists comply with regulatory speed controls, far outweighing traditional demographic characteristics. Using an interpretable machine learning methodology applied to driver behavior data from Kerman Province, Iran, researchers analyzed how individuals respond to three conventional traffic management tools: posted speed limit signs, monetary fines, and demerit point penalty structures. The findings demonstrate a distinct divergence in predictability across these mechanisms, revealing that adherence to physical speed limit signage is governed by relatively direct, systematic behavioral patterns that linear models can capture with moderate accuracy, whereas responsiveness to punitive legal sanctions exhibits substantial behavioral heterogeneity and remains difficult to predict. Across all three regulatory tools, however, self-reported histories of severe speeding, mobile phone distraction, and crash involvement systematically corresponded to diminished compliance sensitivity, underscoring that habitual risk-taking strongly shapes how drivers interact with roadway enforcement systems.
Speed management remains a cornerstone of international traffic safety initiatives, functioning as an indispensable mechanism within the safe system approach to minimize fatal and serious road trauma. Establishing and regulating speed thresholds serves not only to cap maximum travel velocities across highway and urban networks, but also to stabilize vehicular flow and suppress speed dispersion, a dynamic directly associated with elevated collision probabilities. Because the physical energy generated during an impact scales quadratically with vehicular speed, even minor increments in average travel pace produce disproportionate spikes in crash severity and mortality rates. Empirical transportation research consistently indicates that dropping average traffic speeds by as little as one to two kilometers per hour yields meaningful decreases in severe injury collisions. Despite the universal implementation of regulatory speed limits to mitigate the substantial economic, healthcare, and human losses generated by traffic crashes, empirical understanding of how foundational regulatory instruments influence individual driver behavior across distinct motorist profiles has remained underexplored.
While contemporary traffic research frequently evaluates high-profile enforcement technologies such as automated speed cameras and roadside police patrols, foundational administrative instruments like roadside signage, economic penalties, and license point deductions have received comparatively less systematic behavioral evaluation. Posted speed limit signage provides the primary visual notification of legal operating thresholds and theoretically provides general deterrence, yet observational studies indicate that signage effects decay over time without consistent reinforcement. Meanwhile, punitive measures like monetary fines offer immediate economic consequences, and demerit systems introduce cumulative administrative deterrence by recording infractions longitudinally. Existing transportation literature shows that motorist responses to these deterrents vary substantially across demographic cohorts; younger drivers and male motorists historically demonstrate greater resistance to deterrent threats, whereas education and familial responsibilities often encourage cautious driving. Nevertheless, prior investigations predominantly relied on conventional linear statistical models that aggregate violation counts and struggle to isolate nonlinear dynamics or complex interactions between distinct violation histories and socio-demographic features.
To examine these behavioral interactions without relying on rigid linear assumptions, the research team gathered cross-sectional data using a structured questionnaire administered to motorists residing in Kerman Province between August and November 2024. The survey instrument captured detailed information on socio-demographic backgrounds, driving exposure, crash records over the preceding three years, and specific violation typologies, including mobile phone use and dangerous maneuvers. To prevent underrepresentation of specific demographic groups, survey links were disseminated iteratively across regional media channels and online community groups, with subsequent invitations strategically targeted toward cohorts such as female drivers or specific age brackets. The distribution effort yielded an approximate response rate of 40 percent based on invitations directly issued by the investigators. From an initial pool of 150 submissions, screening and data cleaning removed incomplete or inconsistent records, yielding a final analytical sample of 136 valid observations, exceeding the minimum sample size of 96 calculated using the Johnson and Wichern formula.
The analytical framework established three distinct binary dependent variables representing driver compliance sensitivity to each regulatory instrument, designated as speed sign deterrence, monetary fine deterrence, and demerit point deterrence. In the survey instrument, researchers captured these target behaviors through three explicit binary questions inquiring whether motorists complied with posted speed limit signs while operating a vehicle, whether financial fines deterred them from exceeding speed thresholds, and whether administrative demerit points deterred them from speeding. Positive affirmative responses were coded as compliance sensitivity, while negative responses represented non-compliance or insensitivity to the deterrent mechanism. To predict these binary outcomes, the investigation incorporated an extensive set of explanatory predictors consisting predominantly of categorical variables. These features encompassed demographic attributes like age, gender, educational attainment, and marital status, alongside driving exposure metrics such as average daily driving hours, cumulative years of driving experience, prior crash records, vehicle characteristics, and fine-grained violation categories covering excessive speeding, mobile phone distraction, improper overtaking, illegal turns, and equipment defects.
To identify patterns within the survey dataset, the investigators deployed seventeen supervised machine learning algorithms categorized into tree-based ensembles, linear classifiers, and complementary paradigms. The tree-based ensembles included decision trees, random forests, extra trees, gradient boosting machines, extreme gradient boosting, CatBoost, and histogram-based gradient boosting classifiers, all processing ordinal variables unscaled due to decision trees’ scale invariance. The linear modeling category evaluated regularized algorithms including L1-penalized, L2-penalized, and elastic net logistic regression, applying standard scaling to ordinal variables to optimize numerical convergence and feature weighting. The candidate pool was further rounded out by distance-based algorithms, kernel-based classifiers, probabilistic formulations, generalized additive models, and artificial neural networks. Candidate algorithms were evaluated using a unified protocol where the optimal model for each target was selected based on the harmonic mean of out-of-sample test accuracy and the area under the receiver operating characteristic curve, alongside cross-validation stability to ensure generalizability.
To establish robust predictive estimates and prevent data leakage given the modest sample size, researchers implemented an eighty-twenty train-test split combined with five-fold cross-validation on training data. To address class imbalance across the dependent variables, the team utilized SMOTEN, a synthetic minority oversampling technique adapted for categorical feature spaces. SMOTEN synthesizes artificial records by interpolating between minority class instances and their nearest neighbors rather than merely duplicating entries, thereby establishing more balanced decision boundaries. Crucially, oversampling was applied exclusively to training data and executed independently within each cross-validation fold, ensuring validation folds and the terminal holdout test set remained completely untouched by synthetic observations. The full experimental pipeline sequentially partitioned the dataset, applied SMOTEN to the training set, conducted five-fold cross-validation to tune hyperparameters, trained the final candidate model on the oversampled training data, and evaluated generalization on the unseen test set.
Model interpretability was achieved by integrating SHAP, a cooperative game-theoretic method that calculates Shapley values to quantify the contribution of each feature to individual and global predictions. For speed sign compliance, Elastic Net Logistic Regression delivered the most balanced performance, achieving a test accuracy of 0.71, a test precision of 0.80, a test recall of 0.71, a test F1-score of 0.75, and a test ROC-AUC of 0.72. The model demonstrated strong stability, recording a training accuracy of 0.76 and a mean cross-validation accuracy of 0.67 with a standard deviation of 0.07. SHAP analysis demonstrated that severe speeding exceeding 50 kilometers per hour, mobile phone distraction, improper overtaking, illegal U-turns, crash records, and extended daily driving hours contributed toward predicted non-compliance, whereas educational attainment, driving experience, and number of children shifted predictions toward greater compliance, confirming that past behavioral infractions exert far greater influence than demographic backgrounds.
Predicting compliance with punitive legal sanctions proved markedly more difficult for the machine learning algorithms. For fine compliance, the Histogram-based Gradient Boosting Classifier was the most competitive model, attaining a test accuracy of 0.61, precision of 0.38, recall of 0.33, F1-score of 0.35, and ROC-AUC of 0.54, alongside a mean cross-validation accuracy of 0.67 and ROC-AUC of 0.61. For demerit point compliance, Histogram-based Gradient Boosting similarly outperformed alternative algorithms, recording a test accuracy of 0.64, precision of 0.43, recall of 0.33, F1-score of 0.38, and ROC-AUC of 0.56, with mean cross-validation accuracy of 0.61 and ROC-AUC of 0.53. In both tasks, the modest ROC-AUC scores remained only marginally above random classification, illustrating that punitive deterrence sensitivity possesses weak separability when modeled solely from observable driver characteristics, with mobile phone use, prior crash history, and vehicle defects emerging as the primary predictors of sanction insensitivity.
The marked contrast in predictability between physical signage and punitive penalties provides important theoretical, managerial, and practical insights while highlighting notable research boundaries. Roadside speed limit signs provide continuous, direct visual cues that facilitate immediate self-regulation, which linear models can capture effectively, whereas monetary fines and demerit points represent delayed, probabilistic administrative penalties whose deterrent efficacy depends heavily on unmeasured cognitive constructs, such as perceived apprehension certainty, procedural fairness, and punishment severity. These findings suggest that traffic management authorities should transition from demographic targeting toward behavioral risk profiling, utilizing machine learning insights to guide risk-based patrol allocations, targeted educational campaigns for habitual mobile phone users, and refined penalty calibrations. However, because the study relied on a regional cross-sectional sample of 136 motorists from Kerman Province and self-reported questionnaire data subject to recall and social desirability biases, the authors emphasize that these results represent exploratory associations requiring validation through longitudinal tracking and objective telematics rather than deterministic proof of enforcement effectiveness.
Subject of Research: Machine learning evaluation of driver compliance with speed signs, fines, and demerit points
Article Title: Interpretable machine learning for evaluating speed management strategies using driver behavior data, case study of Kerman, Iran
Article References: Khorshidi, N., Nadimi, N., & Shakoori, S. (2026). Interpretable machine learning for evaluating speed management strategies using driver behavior data, case study of Kerman, Iran. Discover Artificial Intelligence, 6(1), Article 1789. https://doi.org/10.1007/s44163-026-02718-7
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02718-7
Keywords: Speed management, Driver compliance, Interpretable machine learning, SHAP analysis, Elastic Net Logistic Regression, Histogram-based Gradient Boosting, Traffic enforcement, Speed limit signage, Monetary fines, Demerit point systems, Risky driving behavior, Mobile phone distraction, Kerman Province, SMOTEN oversampling, Road traffic safety, Deterrence theory, Cross-validation
News Source: Teresa Odom. (November 14, 2026). Machine Learning Links Past Driving Violations to Reduced Traffic Enforcement Compliance. Scienmag.



