Quadratic memory-augmented spatio-temporal transformer graph convolutional recurrent network for traffic forecasting

Traffic forecasting, a core technology within Intelligent Transportation Systems, holds broad application prospects due to its ability to accurately predict future traffic states through the modeling and analysis of complex spatio-temporal traffic data. Nevertheless, due to the complex temporal and spatial heterogeneity of traffic sequences, existing models are difficult to effectively solve the non-stationary problems caused by emergencies. To this end, this paper proposes a Quadratic Memory-Augmented Spatio-Temporal Transformer Graph Recurrent Network (QMAGRN) model based on an encoder-decoder framework. The model consists of three parts: a spatio-temporal Transformer encoder, a quadratic memory-augmented (QMA) module, and a graph convolutional recurrent neural network (GCRU) decoder. Specifically, the spatio-temporal transformer encoder captures the complex spatio-temporal dependencies in traffic data. We designed the QMA module to dynamically update its memory based on incoming data, enabling it to adapt to changing patterns and trends. The QMA module queries the feature information of the memory module on the time and space axis and uses the attention weighting method to perform feature fusion, thereby enhancing the encoder's ability to capture complex spatio-temporal information. This allows the model to maintain information from earlier periods and provide context that helps understand long-term trends and changes, thereby addressing the non-stationarity of traffic data. The GCRU decoder utilizes the features generated by the QMA module as input for its recurrent units. The graph convolutional layers amalgamate historical information from neighboring nodes, thereby enhancing the spatial consistency of predictions. We conducted extensive experiments on five real datasets, and the results demonstrate that our model has achieved state-of-the-art performance. Furthermore, visualizing the learned QMA module enhances the interpretability of the model. Our code and data are accessible via this link: https://anonymous.4open.science/r/QMAGRN-08CD