CWGAN: A Graph Vector Based Traffic Missing Data Adversarial Generation Approach

Traffic speed prediction is an important and basic application in intelligent transportation system. But due to the equipment failure, the sampled time-series data are often corrupted, which induces data missing problems and creates difficulties in traffic speed prediction. In this paper, To solve a special data missing case, successive long-term data missing, we propose a novel approach combining graph vectorization algorithm and conditional Wasserstein generative adversarial nets (CWGAN). Different from traditional generative adversarial nets, CWGAN introduces graph vectors generating from graph vectorization algorithm as its extra conditional input. Apart from this, we implement enough experiments based on various deep learning methods to compare performances. Finally, we find that CWGAN performs better than other comparative methods when data missing is serious.