Multidimensional Behaviour-Based Clustering and Gradient Boosted Regression Trees Delay Model for Three-Wheelers at Un-Signalized Intersections

Abstract

Urban transportation systems in developing countries, particularly India, face significant challenges due to rapid urbanization, heterogeneous traffic conditions, and infrastructural inconsistencies. Unsignalized intersections serve as critical choke points where delay and congestion are exacerbated by driver behaviour, particularly that of three-wheelers (3Ws). This study incorporates geometric features of Un-signalized intersections and behavioral aspect of drivers with the help of trajectory analysis. A structured, multi-dimensional analysis with Hierarchical Agglomerative Clustering (HAC) employed to classify 14 un-signalized intersections located across four Indian cities. These 14 intersections are categorised into three risk levels, with 5 intersections labelled as Moderate-Risk, 4 intersections as High-Risk, and 5 intersections as Low- Risk. This parameter Risk level is developed based upon average delay at intersections, conflict points, aggressiveness score and violation frequency of three-wheelers. Gradient Boosted Regression Trees (GBRT) technique is incorporated to establish a Threewheeler delay prediction model using behavioral aspects of 3Ws and spatial components of Intersections. The prediction model achieved an R2 score of 0.84 and RMSE of 3.2 seconds, indicating strong predictive performance. Feature importance analysis revealed that driver aggressiveness scores, number of conflict points, and intersection visibility are the top three predictors of this delay model. The results validate the hypothesis that incorporating behavioural and spatial data significantly enhances delay prediction accuracy in mixed traffic environments.