DMIR实验室

Abstract:

Domain adaptation on time series data is an important butchallenging task. Most of the existing works in this area arebased on the learning of the domain-invariant representationof the data with the help of restrictions like MMD. How-ever, such extraction of the domain-invariant representationis a non-trivial task for time series data, due to the complexdependence among the timestamps. In detail, in the fully de-pendent time series, a small change of the time lags or theoffsets may lead to difficulty in the domain invariant extrac-tion. Fortunately, the stability of the causality inspired us toexplore the domain invariant structure of the data. To reducethe difficulty in the discovery of causal structure, we relax itto the sparse associative structure and propose a novel sparseassociative structure alignment model for domain adaptation.First, we generate the segment set to exclude the obstacle ofoffsets. Second, the intra-variables and inter-variables sparseattention mechanisms are devised to extract associative struc-ture time-series data with considering time lags. Finally, theassociative structure alignment is used to guide the transfer ofknowledge from the source domain to the target one. Exper-imental studies not only verify the good performance of ourmethods on three real-world datasets but also provide someinsightful discoveries on the transferred knowledge