Simultaneous multimodal demand imputation and forecasting via graph-guided generative and adversarial network

Abstract

Accurate prediction of multimodal transport demand is essential for effective transport planning and management and enables service optimization based on historical and (predicted) future demand. However, dealing with missing data remains a common challenge in multimodal demand analytics. Furthermore, the potential benefits of knowledge sharing across different modes for simultaneous imputation and forecasting have not been thoroughly explored. This study introduces the Graph-guided Generative-Adversarial Imputation and Forecasting Network (GIF) to tackle these challenges. GIF utilizes a Generative Adversarial Network with a generator and a discriminator. The generator simultaneously fills in missing values and predicts future demand, whereas the discriminator differentiates between synthetic and real data. An Encoder-Decoder framework is employed to reconstruct the generated data to the original data to ensure that the important information is preserved. Spatiotemporal features of each mode’s demand are captured via Transformer-encoder layers, whereas the knowledge sharing among multiple modes is facilitated by graph-guided feature fusion of different modes. The proposed method is evaluated on three real-world transport data sets, demonstrating its potential to address forecasting tasks with missing data in multimodal transport systems. Overall, this study provides insights into the effectiveness of cross-modal knowledge sharing and joint imputation and prediction in enhancing the accuracy of multimodal demand prediction.