Abstract:Traffic volume data at fixed detectors are of great importance to track the time-varying states of urban traffic, and the volume of each turning movement (i.e., continuing straight, turning left, turning right) in an intersection plays a significant role in traffic control. It is a difficult issue to determine the complete road network volume due to the limited coverage of detector deployment. Existing traffic volume imputation methods seem to be hardly tractable to address this persistent missing data problem of fixed detectors because they mainly function by utilizing spatiotemporal patterns of historical data to address the problem of intermittent missing traffic data collected from probe vehicles. Considering the unattainable temporal information of the target movement, this paper aims to develop an end-to-end graph neural network-based imputation approach to impute traffic volume data with persistent missing patterns. We regard the road network as a directed graph with each turning movement as a node and employ a meta-learning-based graph attention mechanism with graph embedding in multiple spatial granularities to fully extract the relationship between the target node and its neighbours. In addition, a multitask learning mechanism is designed to improve the accuracy of the imputation results by setting auxiliary tasks according to the volume conservation constraints. The proposed approach has been validated with extensive simulation experiments and applied to field tests for Shenzhen and Shanghai, China, with considerable accuracy and reliability, satisfying the demand for traffic control. To the best of our knowledge, this is the first time that traffic volume has been imputed for persistent missing patterns with a graph neural network-based method.

A Novel Traffic Flow Data Imputation Method for Traffic State Identification and Prediction Based on Spatio-Temporal Transportation Big Data

A Hybrid Data-Driven Framework for Spatiotemporal Traffic Flow Data Imputation

Traffic Data Imputation and Prediction: an Efficient Realization of Deep Learning

An Informer-based Spatio-Temporal Network for Traffic Flow Forecasting

A Transfer Learning–Based LSTM for Traffic Flow Prediction with Missing Data

A Latent-Factor-Model-Based Approach for Traffic Data Imputation with Road Network Information

Spatial-Temporal Congestion Identification Based on Time Series Similarity Considering Missing Data

A Kriging based spatiotemporal approach for traffic volume data imputation

Urban Traffic Flow Forecasting Based on Memory Time-Series Network.

A Deep Learning Framework for Traffic Data Imputation Considering Spatiotemporal Dependencies

Missing Data Imputation for Traffic Flow Based on Combination of Fuzzy Neural Network and Rough Set Theory.

Missing Traffic Data: Comparison of Imputation Methods

An Experimental Evaluation of Imputation Models for Spatial-Temporal Traffic Data

Bidirectional Spatial-Temporal Traffic Data Imputation via Graph Attention Recurrent Neural Network

A hybrid model for missing traffic flow data imputation based on clustering and attention mechanism optimizing LSTM and AdaBoost

Traffic Prediction with Transfer Learning: A Mutual Information-based Approach

Big Data-Driven Based Real-Time Traffic Flow State Identification and Prediction

Traffic Prediction With Missing Data: A Multi-Task Learning Approach

Turning traffic volume imputation for persistent missing patterns with GNNs

An Intelligent Deep Learning Framework for Traffic Flow Imputation and Short-term Prediction Based on Dynamic Features

Traffic Status Prediction Based on Multidimensional Feature Matching and 2nd-Order Hidden Markov Model (HMM)