Hyperparameter Tuning to Optimize Implementations of Denoising Autoencoders for Imputation of Missing Spatio-temporal Data

Abstract Spatio-temporal data collected from sensors can sometimes have gaps where data is missing. Transportation planners and engineers use such data to perform various different types of analyses, but the gaps in the data make it difficult to make accurate predictions. Denoising Autoencoders are used to generate a cleaner version of the input signals. This includes generating data for missing or dropped input signals as well. We used three different deep learning models to implement Denoising Autoencoders to measure each model’s accuracy. Since tuning the hyperparameters can influence the accuracy of the predictions of a model, each model was optimized by focusing on four main parameters. Each of the hyperparameters had a different effect on each of the models. The simplest of the models yielded much better results than anticipated after optimizing. The most complex of the models was still better, but only slightly. That slight improvement required one hundred times the computational cost required by the simplest model. Tuning the hyperparameters to optimize a simpler model can prove more beneficial than creating a more complex model that is slower and difficult to optimize.