Parameter-Transferred Irreducible LSTM for Traffic Data Imputation

We propose an imputation algorithm for missing spatiotemporal data based on long short-term memory (LSTM) model factorization in a traffic environment where the roadside units (RSUs) collect traffic speed data. We considered a scenario where data collection by RSUs occurs for each road segment, but the absence of RSUs results in an incomplete dataset. To enhance imputation accuracy, we mitigate the risk of error propagation of model training on the entire dataset and take into account the spatiotemporal correlation of the dataset. The proposed algorithm can reduce the dimensionality of the input dataset by employing an adjacency matrix to identify data both highly correlated and connected to the target road segment, subsequently transforming parallel datasets into a serial format. Then, we extrapolate the missing data using an irreducible LSTM model, which is a factorization of a standard LSTM model. To enhance imputation performance, we also adopt spatial interpolation on extrapolated data across multiple paths that lead to the target road segment. Extensive experiment results using synthetic and real-world datasets confirm that the proposed algorithm outperforms other imputation algorithms in terms of imputation accuracy measured by the root mean square error (RMSE) and the mean absolute error (MAE) as well as the space complexity measured by the number of model parameters. In particular, the experiments with real-world datasets show that the proposed algorithm consistently achieves high imputation accuracy across a wide range of traffic scenarios, including actual traffic congestion and rapid traffic fluctuations.