Is Classical LSTM more Efficient than Modern GCN Approaches in the Context of Traffic Forecasting?

Traffic forecasting is one of the most difficult challenges in the area of ITS (intelligent transportation systems) because of complex spatial correlations on road networks and non-linear temporal dynamics of changing road conditions. To address these issues, researchers proposed models that combine GCNs (Graph Convolution Networks) and RNNs (Recurrent Neural Networks), in order to inherit the advantages of both of them and become capable of extracting spatiotemporal correlations. Restricting the efficiency of the models by their precision without concern for their structure made the models become more complex, although simple models sometimes produce better results. In this research, we introduce a simple model, called Long Short-Term Memory network for Traffic Forecasting (LSTM-TF), which uses the LSTM for extracting spatial-temporal dependencies. Experiments show that the LSTM-TF outperforms state-of-the-art baselines on real-world traffic datasets, proving our hypothesis that simple models as the LSTM-TF produce sometimes better results than more complex ones.