Aggregating From Multiple Target-Shifted Sources
This work addresses a critical bottleneck in multi-source domain adaptation for scenarios with target-shifted sources, offering a unified solution for practitioners dealing with limited or no target labels.
The paper tackles the problem of multi-source domain adaptation when source domains have different label distributions, proposing a method that aggregates sources based on semantic conditional similarity rather than marginal distribution. The results show significant empirical improvements over baselines in experiments across three domain adaptation scenarios.
Multi-source domain adaptation aims at leveraging the knowledge from multiple tasks for predicting a related target domain. Hence, a crucial aspect is to properly combine different sources based on their relations. In this paper, we analyzed the problem for aggregating source domains with different label distributions, where most recent source selection approaches fail. Our proposed algorithm differs from previous approaches in two key ways: the model aggregates multiple sources mainly through the similarity of semantic conditional distribution rather than marginal distribution; the model proposes a \emph{unified} framework to select relevant sources for three popular scenarios, i.e., domain adaptation with limited label on target domain, unsupervised domain adaptation and label partial unsupervised domain adaption. We evaluate the proposed method through extensive experiments. The empirical results significantly outperform the baselines.