A Data-Level Augmentation Framework for Time Series Forecasting With Ambiguously Related Source Data

Rui Ye, Qun Dai

Abstract: Many practical time series forecasting (TSF) tasks are plagued by data limitations. To alleviate this challenge, we design a data-level augmentation framework. It involves a time series generation (TSG) module and a source data selection (Sel-src) module. TSG aims to achieve better generation results by considering both the global profile and temporal dynamics of series. However, when only few target data is available, TSG module may tend to simulate the limited target samples, leading to poor generalization performance. A natural idea for this problem is to seek help from related source domain, which can provide additional useful information for TSG module. Here we consider a more complex situation, where the relevance between source and target domains is ambiguous. That is, irrelevant samples may exist in the source domain. Blindly using all the source data may lead to counterproductive results. To meet this challenge, Sel-src module is designed to select effective source samples by Inter-Representation Learning (Inter-RL) and Intra-Representation Learning (Intra-RL). Effectiveness of this algorithm is underpinned from two aspects: the quality of the augmented data and the accuracy improvement upon the augmentation.