Deep Regression Modeling for Imbalanced and Incomplete Time-Series Data

During the collection of time-series data, many reasons lead to imbalanced and incomplete datasets. Consequently, it becomes challenging to develop deep convolutional models without suffering from overfitting. Our objective in this paper was to investigate an emerging but rather underutilized framework of Conditional Generative Adversarial Networks (cGANs) for improving deep regression models for time-series data with an imbalanced and incomplete distribution. First, we investigated the potential of using a vanilla cGAN as a data imputation to improve the generalizability of the developed models to unseen data in such datasets. Next, we proposed a modified cGAN architecture with improved extrapolation and generalizability of the regression models. Our investigations used an imbalanced synthetic non-stationary dataset, a real-world dataset in Parkinson's disease (PD) application domain, and one publicly-available dataset for Negative Affect (NA) estimation. We found that vanilla cGAN failed to generate realistic time-series data due to severe mode collapse, limiting its application as a data imputation for imbalanced and incomplete data. Importantly, the proposed cGAN framework significantly improved extrapolation and generalizability for the prediction of regression scores with an average improvement of 56%, 34%, and 18%, respectively, in mean absolute error for the synthetic, PD, and NA datasets when compared with traditional Convolutional Neural Networks. The codes are publicly available on Github.