Assessment of bus speed influencing factors through the exploitation of machine learning techniques

Abstract

Bus commercial speed is among the most influencing parameters of public transport quality of service. Literature has highlighted many different factors that affect bus speed in different ways. This paper focuses on the development of predictive models for the estimation of bus commercial speed, as well as on the determination of the effect of several variables on bus speed between successive stops. For the need of predicting bus commercial speed, three (3) machine learning models were developed; a Support Vector Regression (SVR); a Random Forest Regression (RFR); and an Artificial Neural Network (ANN). The models were based on data referring to 10 lines of the bus network of Thessaloniki, Greece. Data included several features such as bus dwell time, passenger demand, bus stop location, bus stop infrastructure etc. Results indicate a satisfactory performance by all models, while the best accuracy was achieved from the ANN model. Afterwards, the analysis focused on the determination of the effect of the explanatory variables on commercial bus speed. To achieve that, the Shapley Additive Explanations (SHAP) algorithm was applied, based on the ANN model. Bus stop spacing was identified as the most contributing feature to the model's output; distance of more than 450 meters between 2 successive stops was found to have a positive effect on bus speed. Single-shelter bus stops were associated with higher bus commercial speed, while the location of bus stops in the middle of a building block proved to affect bus speed in a positive way. Dwell time was also a crucial parameter, having a decreasing effect on bus speed, as its value increases. Additionally, it was found that an increase in the number of traffic lights and turns along a route section, affects bus speed negatively. Finally, no apparent relation was observed between total passenger traffic and bus speed.