TrajTransGCN: Enhancing trajectory prediction by fusing transformer and graph neural networks

Abstract

This paper proposes a novel model named TrajTransGCN for taxi trajectory prediction, which leverages the power of both graph convolutional networks (GCNs) and Transformer. TrajTransGCN first passes the input through the GCN layer and then combines the GCN outputs with one-hot encoded categorical features as input to the transformer layer. This paper evaluates. TrajTransGCN uses real-world taxi trajectory datasets in Porto and compares it against several baselines. The experimental results show that TrajTransGCN outperforms all the other models in terms of both RMSE and MAPE. Specifically, the model achieves an RMSE of 0.0247 and a MAPE of 0.09%, which are significantly lower than those of the other models. The results demonstrate the effectiveness of the proposed model in predicting taxi trajectories, indicating the potential of leveraging both GCN and transformer layers in trajectory prediction tasks. In addition, this paper includes ablation experiments to demonstrate the effectiveness of using one-hot encodings of classification labels in complex real-time scenarios. In addition, a parameter study is carried out to examine how the TrajTransGCN's performance is impacted by the learning rate, the quantity of Transformer layers, and the size of the hidden dimension of the Transformer layer.