Enhancing Travel Time Prediction in African Public Transit Systems:A Comparative Study of Machine Learning and Deep LearningApproaches

While accurate travel time estimation remains a challenge in African public transit systems, especially in under-resourced cities like Bamako, this study explores the use of machine learning and deep learning models to predict short-term travel times between consecutive bus stops. Existing models often treat transport data as static and tabular, failing to capture the complex spatial and temporal dynamics inherent in urban bus systems. This study evaluates and compares conventional models (XGBoost, Random Forest, Gradient Boosting) with a hybrid Graph Neural Network and Long Short-Term Memory (GNN+LSTM) architecture. Using preprocessed General Transit Feed Specification (GTFS) data, the study engineered spatial, temporal, and operational features, including haversine distance between consecutive stops, which exhibited a correlation of approximately 0.88 with traveltime. The Graph Neural Network (GNN) component, implemented using Spektral Graph Convolutional Network (GCN) layers on a real transit graph adjacency matrix, encoded spatial relationships among bus stops through message-passing, while the LSTM captured temporal patterns across trip sequences. With the inclusion of the distance feature, XGBoost achieved the best overall performance with a Mean Absolute Error (MAE) of 15.24s, Root Mean Squared Error (RMSE) of 56.51s, and 𝑅2 of 0.7970. The end-to-end GNN+LSTM model achieved a MAE of 26.34s, RMSE of 75.80s, and 𝑅2 of 0.6348, demonstrating the value of jointly modeling spatial graph structure and temporal trip dynamics. The results highlight that inter-stop distance is the single most predictive feature for travel time estimation, and that spatio-temporal deep learning architectures offer competitive performance with potential advantages on larger, more complex transit networks.