Abstract
Machine learning provides an effective approach for modelling discharge headway at signalized intersections under mixed traffic conditions, where conventional analytical models based on homogeneous assumptions are often inadequate. This study presents a comparative evaluation of Linear Regression (LR), Support Vector Regression (SVR), and Random Forest (RF) models for predicting discharge headway at internal (CBD) and external (non-CBD) intersections in Srinagar, India. Field data were collected using video-based surveys at six signalized intersections, resulting in 1085 observations for internal intersections and 498 for external intersections. The dataset was randomly split into 80% training and 20% testing sets, and model robustness was evaluated using five-fold cross validation. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), coefficient of determination (R²). Results show that Random Forest outperforms LR and SVR in both environments. At internal intersections, RF achieved MAE = 0.356 s, RMSE = 0.519 s, and R² = 0.705, while at external intersections, it achieved MAE = 0.488 s, RMSE = 0.672 s, and R² = 0.742. The external Random Forest model showed good predictive ability on the test set, but the variability in the cross-validation results was comparatively larger than that observed for the internal dataset. Thus, the reported performance should be interpreted within the context of traffic environments comparable to the non-CBD intersections investigated in this study, and larger external datasets are expected to further enhance the robustness and generalizability of the developed models. The Random Forest model achieved a 65.6% reduction in MAE for internal intersections and a 72.7% reduction in MAE for external intersections compared with the HCM/IHCM benchmark (1.920 s), demonstrating a significant improvement over the conventional capacity estimation method. Feature importance analysis indicates that vehicle length has the highest predictive importance at internal intersections, whereas green time and distance to median show greater predictive importance at external intersections. These findings demonstrate the effectiveness of Random Forest and indicate that traffic environment can influence model performance and predictor importance, with implications for capacity estimation and signal design under heterogeneous traffic conditions.