Robust spatial temporal imputation based on spatio-temporal generative adversarial nets

Abstract

Spatial temporal imputation is an important part of intelligent transportation system, and spatial temporal data missing also largely affects the performance of its downstream tasks. Spatial temporal data are affected by spatial, temporal, and unexpected conditions, and has a complex change pattern. Existing researches mostly focus on time series and often ignore the modeling of spatial dependency. Moreover, the imputation performance is poor under high missing rate or block missing scenarios, needing to enhance the model robustness. Therefore, we proposed a spatio-temporal generative adversarial imputation network (named STGAIN) to handle with the complex spatial temporal dependency and the diversity of missing conditions on spatio-temporal data. It includes a spatio-temporal generator and a spatio-temporal discriminator, which uses a graph convolutional neural network as a spatial aggregator and a one-dimensional deformable convolutional neural network as a temporal extractor. The generator observes the unmissing part data and the spatial relationships to impute the missing parts. The discriminator then takes the completed data as input and attempts to determine which components were actually observed and which were imputed. The spatio-temporal modeling ability was improved in the competition with each other. In order to make the generator in GAN distinguish the observed part from the missing part, a fusion loss function (including reconstruction loss and BCE loss) was designed for the generator training. Finally, experiments were conducted on two real highway datasets for different missing rates. And we test the model robustness under different size of dataset and noise interference. The results show that STGAIN significantly outperforms the existing methods.