Unified Principles For Multi-Source Transfer Learning Under Label Shifts
We study the label shift problem in multi-source transfer learning and derive new generic principles to control the target generalization risk. Our theoretical grounded framework unifies the principles of conditional feature alignment, label distribution ratio estimation and domain relation weights estimation. Moreover, the inspired practice provides a unified algorithm for various multi-source label shift transfer scenarios: learning with limited target data, unsupervised domain adaptation and label partial unsupervised domain adaptation. We evaluate the proposed method on these scenarios by extensive experiments and show that our proposed algorithm can significantly outperform the baselines.
PDF Abstract