Improving time series forecasting with mixup data augmentation

Mixup is a domain-agnostic approach for data augmentation, originally proposed for training Deep Neural Networks (DNNs) for image classification. It obtains additional data for training by sampling from linear interpolations of model inputs and their labels. While proven to be effective for computer vision (CV) and natural language processing (NLP) tasks, it remains unknown if mixup can bring performance improvement for DNNs developed for forecasting tasks. We propose several different approaches for applying mixup to train neural forecasters and evaluate them on benchmark datasets. The study highlights the capability of mixup in enhancing model performance across a wide range of hyper-parameter settings in specific datasets, and paves the way for mixup’s potential success in broader time series forecast use cases.