Research On Traffic Data Imputation Methods Based On Generative Adversarial Networks

In recent years,with the rapid development of big data technology and artificial intelligence technology,data-driven intelligent transportation applications bring great convenience to people’s daily life.However,data missing is a common and inevitable problem that seriously affects the correctness and effectiveness of corporate and institutional decision-making.Accurately impute the missing traffic data,on the one hand,it can improve the quality of the data and make it produce more reliable output in data analysis and mining tasks.On the other hand,it can help enterprises and management departments to conduct effective analysis and make reasonable decisions.In real application scenarios,the traffic data imputation task mainly has the following two difficulties:(1)Modeling and learning data distribution are effective ways to implement model imputation,and traffic data also has complex dynamic spatial-temporal correlations in both time and space dimensions.Therefore,how to effectively model the distribution of data,fully mine and utilize these spatial-temporal correlations is a major challenge;(2)The problem of missing labels.The true labels of missing values are often unavailable,which resulting in a lack of effective supervision information for model training and worrying imputation quality.Existing works cannot deal with the above difficulties at the same time,and this paper conducts in-depth research.Firstly,a generative model ST-GAN based on the generative adversarial network is proposed,which interpolates traffic data by simultaneously considering both the spatial and temporal characteristics of traffic data.In the temporal dimension,a multi-head self-attention mechanism that can perceive the changing trend of time series is used to capture the dynamic correlation of traffic data.In the spatial dimension,a dynamic graph convolution method is used to adaptively capture the correlation of the traffic data in the spatial dimension.Then the framework of the generative adversarial network is innovatively adopted to realize the imputation of the spatial-temporal traffic data.Secondly,in order to further improve the imputation accuracy and promote the application of the model in practical scenarios,this paper further proposes a spatial-temporal generative adversarial network model SST-GAN based on the ST-GAN model,which integrates the self-supervised learning.Specifically,a generative self-supervised learning method is introduced,which takes the self-attributes of non-missing positions as the self-supervised label so as to guide the model to perform more effective training,strengthen the generation ability of the generator,and make the generated data closer to the real data.In addition,a feature-based data enhancement module is introduced to perform small-scale masking on the input data,so that the model can mine potential information from more representative data,thereby improving the performance of the model and making it more efficient in the traffic data imputation tasks.Finally,extensive experiments are conducted on the real traffic flow dataset Pe MS.The experimental results show that the ST-GAN and SST-GAN models proposed in this paper significantly improve imputation accuracy compared with the existing baseline models.Both models are able to take full advantages of the spatial-temporal correlations in spatial-temporal traffic data to model the true distribution of missing data,and generate high-quality traffic data.The improved SST-GAN model which adopts the self-supervised training framework with regarding its own attributes as labels,can fully mine and analyze the pattern characteristics of the traffic data itself,thus it further improves the traffic data imputation accuracy.