Application of machine learning models for evaluating the impact of roadside advertising on driver discomfort

• Novel approach to evaluating driver discomfort from roadside advertisements using photometric measurements and machine learning techniques. • Integration of luminance and contrast analysis with automated classification models for objective risk assessment. • Comparison of four machine learning models (logistic regression, decision trees, random forests, neural networks), with neural networks achieving the highest classification accuracy (AUC = 95.32 %). • Identification of key factors influencing discomfort, with luminance intensity and contrast being the most significant contributors. • Potential application in real-time adaptive advertisement control systems, optimizing roadside lighting regulations for improved traffic safety Road safety depends critically on the effect of roadside advertising on driver discomfort since too strong brightness and contrast changes can cause visual strain and lower attention span. This work considers important brightness and road infrastructure factors to test several machine learning models for estimating driver discomfort generated by advertising billboards. Train and evaluate four classification models—logistic regression, decision tree, random forest, and neural network—using a dataset including brightness measurements and road conditions. Accuracy, sensitivity, specificity, and the area under the ROC curve (AUC) were used in evaluation of the models Model performance was optimized using LASSO and ElasticNet regularization methods; model validation was conducted using 10-fold cross-valuation. Outstanding in balancing false positive and false negative rates, the neural network showed 93.18%, sensitivity (96.97%), and specificity (81.82%). With an accuracy of 86.36% logistic regression showed a lower specificity (45.45%), which increased false positive rate. While the decision tree shown the lowest classification capacity (75.00% accuracy, 72.18% AUC), the random forest model fared rather well (79.55% accuracy, 81.13% AUC). Maximum luminance, contrast, and average luminance were the main factors predicting driver discomfort; road infrastructure measures had little effect. Particularly neural networks, machine learning algorithms can evaluate motorist uneasiness about roadside advertising with great accuracy. The results offer information for formulating rules to increase road safety and maximize the circumstances of roadside illumination. Advanced deep learning configurations and real-time discomfort detection systems in dynamic driving settings should be investigated in next studies.