Efficient source-free time series fitting via parameter subspace disentanglement

The growing demand for personalized and private on-device applications highlights the importance of source-free, unsupervised domain adaptation (SFDA) methods, particularly for time series data where individual differences result in large domain shifts. As mobile devices with embedded sensors become ubiquitous, optimizing SFDA methods for parameter utilization and data sampling efficiency in time series contexts becomes critical. Time series personalization is necessary to account for the unique patterns and behaviors of individual users to improve the relevance and accuracy of predictions. In this work, we introduce a new paradigm for source model preparation and target-side fitting that aims to improve both parameter and sample efficiency during the target-side fitting process. Our approach reparameterizes the weights of the source model with Tucker-style decomposed factors during the source model preparation phase. Then, at the time of target-side adjustment, only a subset of these decomposed factors are fine-tuned. This strategy not only improves the parameter efficiency, but also implicitly regulates the adaptation process by limiting the capacity of the model, which is essential for personalization in diverse and dynamic time series environments. Furthermore, the proposed strategy achieves overall model compression and improves inference efficiency, making it ideal for resource-constrained devices. Extensive experiments on various SFDA time series benchmark datasets demonstrate the effectiveness and efficiency of our approach and highlight its potential for further developing personalized on-device time series applications.