Time Series Imputation Using a Neural Network Model
Conference
66th ISI World Statistics Congress
Format: WSC 2027: IPS Abstract
Keywords: deep neural networks, imputation, time-series-models, transfer-learning
Abstract
Sensor data—such as that derived from induction loops and traffic cameras—is a valuable, readily available source for estimating inbound road tourism when integrated with mobile data, that lacks comprehensive passenger register data, unlike the detailed records available for ports, airports, and rail. Another advantage of this data is its high temporal resolution. However, the intermittent malfunction of these sensors frequently introduces significant data gaps into the time series.
Imputation is thus needed to fill up the gaps. For this task, we propose a novel deep learning model for multivariate time series imputation. The architecture integrates Long Short-Term Memory (LSTM) layers to effectively capture temporal dependencies and features, alongside autoencoder-like layers to extract and leverage spatial features from co-located loops and cameras. It is possible to include extra additional regressors too. The training process is sequential: the LSTM and autoencoder components are pre-trained independently and then combined using shallow, fully connected layers which are fine-tuned on the final dataset.
This implementation, built using Python/Keras, employs custom layers and metrics essential for facilitating this multi-stage, sequenced training approach and effectively handling the time-varying nature of the gaps. Furthermore, this general architecture is readily adaptable for application to other multivariate high frequence time series with strong spatial dependencies.