Cyclic Generative Adversarial Networks with KNN-transformers for missing traffic data completion

In the face of the huge amount of intelligent transportation data, it is necessary and important to collect and statistically process it. Due to adverse weather conditions, sensor malfunctions and other reasons, the collected data inevitably contains missing data. Aiming at the phenomenon of missing traffic data, we propose an interpolation method of missing traffic data based on Cyclic Generative Adversarial Networks with hybrid KNN-Transformer method (KT-CyclicGAN). This method effectively utilizes K-Nearest Neighbor as prior knowledge to guide network training, employs transformers to extract the spatiotemporal relationships present in traffic data, and reconstructs missing data. In GANs, it uses a multi weight sharing cyclic structure to thoroughly learn the spatiotemporal sequences in the traffic data, resulting in accurate imputed data. We assess the performance of the proposed model using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R-squared, and the Concordance Correlation Coefficient (CCC), while simulating three missing data scenarios on the PEMSD4 dataset. The experimental results show that the algorithm proposed in this paper can deal with various missing scenarios and missing rates more effectively than the other five algorithms. Even with a high missing rate of up to 90%, the data imputed by KT-CyclicGAN can still fit the real data quite well.