Causal Structure Alignment Based Apporach For Time Series Domain Adaptation

The construction of machine learning models normally requires a mass of labeled data.But the high cost of labeling data and the difficulty of satisfying the basic assumption of independently identically distribution for real-life multi-source data both make it difficult to implement machine learning models in real scenarios.Domain adaptation offers some solutions to this problem,which mainly uses the currently available labeled data to obtain models that can be adapted to the target domain,effectively solving the above two problems and building machine learning models faster and better in complex real-life scenarios.However,current domain adaptation methods are mainly for image data,while domain adaptation for time series data is an equally important and challenging task.Most of the existing work in this field is currently based on constraints such as maximum mean discrepancy or gradient reverse layers to learn domain invariant representations of the data.However,the extraction of domain-invariant representations is very difficult for time series data due to complex dependencies between timestamps in the time series,such as domainspecific offsets and time lags,where tiny differences between domains are passed over the time series.Fortunately,the stability of causality between variables motivates us to explore the domain-invariant structure of the data.To reduce the difficulty of causal structure discovery,we relax it to a sparse associative structure and propose a new time series domain adaptation algorithm based on sparse associative structure alignment.In this paper,we address the unsupervised domain adaptation problem for time series data,and the main research contents and contributions are as follows.(1)A time series domain adaptation method based on associative structure alignment is proposed.The algorithm starts from the domain-invariant structure between time series to avoid the difficulty of directly aligning the feature distributions of different domains.In this paper,we use a two-stage attention mechanism to extract the sparse associative structure between time series and further align the structure of source and target domains by minimizing the maximum mean discrepancy.We weight and reorganize the features based on the aligned sparse associative structure to obtain the domain-invariant features that are finally used for prediction.(2)The domain-specific Offsets and Time Lags problems in time series domain adaptation are solved.The complex dependencies in the time dimension of time series data will increase the time series discrepancies between different domains,and the domain-specific offsets and time lags will also increase the time series discrepancies between different domains,which makes the extraction of domain-invariant representations very difficult.In this paper,we obtain adaptative segment summarization by segmenting the time series,and align the segment weights and structure matrices in the source and target domains to solve the problem of different offsets and time lags in the source and target domains.Comparative experimental results verify the good performance of our methods on four real-world datasets and our model is able to extract domain-invariant information more effectively on time-series data.The visualization results of the associative structure show that the method in this paper can extract sparse domain invariant structures.