Weighted Partial Optimal Transport for Multi-Source Partial Domain Adaptation

We develop a theoretical and algorithmic framework for multi-source partial domain adaptation (MSPDA) by deriving a generalization bound that relates the target loss to weighted empirical source losses and source-specific partial Wasserstein distances. This bound motivates a partial optimal transport algorithm, termed MS-WARMPOT, that shares a common feature extractor across domains, addressing multi-source heterogeneity. MS-WARMPOT learns source-target sample weights that suppress outlier classes and prevent negative transfer, thereby unifying multi-source domain adaptation (MSDA) and partial domain adaptation within a single framework. Experiments on standard MSDA and MSPDA benchmarks demonstrate competitive performance against the state-of-the-art methods.