BOOST: Batch Ordering for Training Deep Forecasting Models

Achieving higher accuracies in training deep learning models while avoiding over-fitting is essential when dealing with difficult-to-collect datasets, particularly system-level metrics datasets. Although switching models can significantly improve the resulting accuracies in many applications, it comes at a higher computational cost. Alternatively, modifying the order of samples in the training set could help the model converge to superior accuracies and prevent the high cost of any modifications in the model architecture. One of the well-studied approaches to achieving this goal, which is widely used in image and video processing tasks but not time series forecasting, is curriculum learning (CL). This rarity is due to the computational costs and mediocre performance of CL methods, particularly in metrics time series datasets. In this paper, we propose a batch-ordering method called BOOST to resolve the limitation of the CL methods when dealing with large time series by incorporating a feature extraction and clustering pipeline. The experimental results of applying BOOST to azure, electricity, traffic, and wiki time series datasets indicate that the mean absolute error improved by 18.27%, 10.33%, 7.13%, and 1.28% when using DeepAR-Gaussian, C2FAR, DLinear, and LogTransformer forecasting methods, respectively.