Traffic Data Imputation with Ensemble Convolutional Autoencoder.

Intelligent transportation systems and related applications rely on high-quality traffic data. However, the data collected in real-world is often incomplete, which compromises the system performance. Traffic data imputation estimates the missing values by analyzing traffic flow features, therefore can improve the performance of related applications. Traditional imputation methods mainly focus on isolated traffic data sensors or road sections and show their limitations in representing complex spatial-temporal features. In this paper, we propose a novel ensemble model named ensemble convolutional autoencoder for the task. The observed values, together with the missing points are reconstructed into a two-dimensional matrix by the extracted spatial-temporal relation. Convolutional and deconvolutional layers are adopted to encode and decode spatial-temporal features, respectively. Besides, we train autoencoders with different input feature maps and ensemble the outputs by linear combination. Experimental results show that compared with other traffic data imputation methods, the proposed method can achieve better accuracy and has stable performance under various missing data scenarios with different types and rates.