Dynamic Graph Convolutional Network-based Long Short-Term Memory for Taxi Trip Time Prediction

In the last few years, traveling time has increased rapidly due to improper management of the time schedule and road traffic which also directly affects taxi users. The existing models do not offer appropriate results for the prediction of time taken to complete a taxi trip due to poor spatial and temporal information of the features. This paper proposes a Dynamic Graph Convolutional Network-Long Short-Term Memory (DyGCN-LSTM) model for the prediction of taxi trip time. The Extreme Gradient Boosting (XGB) model method is used for the analysis of feature importance based on Shapley Additive Explanations (SHAP) value. Mean Absolute Error (MAE), precision, Root Mean Squared Error (RMSE), recall, and f1-score are the metrics taken into consideration for performance evaluation. The DyGCN-LSTM is prone to a low error rate of $\mathbf{7. 0 5}$ for RMSE and $\mathbf{2. 0 4}$ for MAE on the New York City (NYC) taxi trip record-2019 dataset, with minimum error of $\mathbf{5. 2 2}$ for RMSE and $\mathbf{1. 6 5}$ for MAE on the NYC taxi trip record-2022 dataset. This has hence proved to be superior in comparison to the existing Long Short-Term Memory Convolutional Neural Network(LSTM-CNN) autoencoder method.