Bidirectional Imputation of Spatio-Temporal Data based on LSTM with Parameter Transfer

In this paper, we propose a bidirectional imputation algorithm for spatio-temporal traffic speed data based on Long Short-Term Memory (LSTM) architecture with parameter transfer in vehicle to infrastructure (V2I) networks. We consider a scenario in V2I networks where an Road Side Units (RSU) on the road does not operate temporarily and thus the traffic speed data cannot be collected. This makes any services that rely on the traffic speed data at the RSU be unavailable. For uninterrupted and seamless V2I services, an efficient and low complexity data imputation algorithm is imperative. The proposed algorithm is based on the architecture that includes multiple LSTM layers with parameter transfers, which can explicitly take into account the spatio-temporal characteristics of the traffic speed data. By transferring parameters from one LSTM layer to its adjacent LSTM layer, the complexity associated with the algorithm can be significantly reduced. The proposed algorithm includes bidirectional imputation, which can further improve imputation accuracy. Our simulation and experiment results confirm that the time for training and data imputation of the proposed algorithm can be significantly reduced while maintaining imputation accuracy.