A relationship-aligned transfer learning algorithm for time series forecasting

Rui Ye, Qun Dai

Abstract: Though many excellent methods have been designed for time series forecasting, they often entail sufficient training data from the same domain, which might be difficult to fulfill in some real-world applications. To alleviate this problem, a Relationship-Aligned Transfer Learning (aRATL) algorithm is proposed in this paper, in which the transfer learning process is implemented across different datasets. Whereas in some real scenarios, relevance between alternative source datasets and the target dataset may be ambiguous. For addressing this challenge, instead of calculating similarities at the data level, RATL tends to select the source model whose parameters facilitate the completion of the target task. The knowledge transfer in RATL contains representation relationship-aligned and regression relationship-aligned stages. The former one aims to enhance the representation ability of the target model by aligning relationships, in the form of triplets, between source and target models. The latter one aims to borrow regression experiences from the source model. Since predictions obtained by the source regression model are not always precise, RATL stresses the good results, but ignores the bad ones. Effectiveness of this proposed algorithm is underpinned by extensive experiments on five benchmark time series datasets, compared with several other state-of-the-art methods.