Self-supervised generative adversarial learning with conditional cyclical constraints towards missing traffic data imputation

Abstract

Accurate traffic data imputation aims to fill in missing traffic values with observations as much as possible, which has long been a challenging task that affects its exploitation and application in intelligent transportation system. However, most available generative models lack extensive and effective conditional constraints during their training phase. Hence, the repair accuracy and scalability of these models are poor when dealing with complex missing cases. To fill this research gap, we propose a novel framework called self-supervised generative adversarial learning with conditional cyclical constraints (i.e., C3S2-GAL) to address the traffic data imputation problem. C3S2-GAL employs two self-supervised sparse autoencoders (i.e., S2-SAE) that share the same network structure as it's generator and discriminator to perform the fine-grained mapping and precise output of the input partial observations. With the help of multiple loss constraints, S2-SAE learns the shallow hidden relationships between observations and then generates the initial complete data. Moreover, we further enhance the repair capability of S2-SAE by calculating and integrating the pseudo-labels loss of implicit category information within the pre- and post-repair traffic data. Also, we introduce a cyclical consistency constraint to provide reliable supervision for observed and unobserved traffic values and adopt an iterative learning strategy to ensure that C3S2-GAL yields high-quality results. Finally, extensive experiments are conducted on the publicly available Guangzhou and Seattle-Loop traffic speed datasets to evaluate C3S2-GAL under different missing scenarios and missing rates. The comparison results, convergence analysis, sensitivity analysis and ablation study confirm the superiority of C3S2-GAL over existing state-of-the-art baseline models.