Estimation of missing values in heterogeneous traffic data: Application of multimodal deep learning model

Abstract

With the development of sensing technology, a large amount of heterogeneous traffic data can be collected. However, the raw data often contain corrupted or missing values, which need to be imputed to aid traffic condition monitoring and the assessment of the system performance. Several existing studies have reported imputation models used to impute the missing values, and most of these models aimed to capture the spatial or temporal dependencies. However, the dependencies of the heterogeneous data were ignored. To this end, we propose a multimodal deep learning model to enable heterogeneous traffic data imputation. The model involves the use of two parallel stacked autoencoders that can simultaneously consider the spatial and temporal dependencies. In addition, a latent feature fusion layer is developed to capture the dependencies of the heterogeneous traffic data. To train the proposed imputation model, a hierarchical training method is introduced. Using a real world dataset, the performance of the proposed model is evaluated and compared with that of several widely used temporal imputation models, spatial imputation models, and spatial–temporal imputation models. The experimental and evaluation results indicate that the values of the evaluation criteria of the proposed model are smaller, indicating a better performance. The results also show that the proposed model can accurately impute the continuously missing data. Furthermore, the sensitivity of the parameters used in the proposed deep multimodal deep learning model is investigated. This study clearly demonstrates the effectiveness of deep learning for heterogeneous traffic data synthesis and missing data imputation. The dependencies of the heterogeneous traffic data should be considered in future studies to improve the performance of the imputation model.