SADA: Self-Adaptive Domain Adaptation From Black-Box Predictors

Domain adaptation from black-box predictors aims to perform domain adaptation using predictions from a model trained on the source domain, thereby avoiding privacy issues associated with source domain data and enabling the training of a portable lightweight model on the target domain. Due to domain shift, predictions from the source domain model often contain noise and accumulate confirmation bias. Past works addressing confirmation bias typically employ model-agnostic methods, overlooking the learning status of the model, thereby resulting in issues such as class imbalance and training data scarcity. To address these two issues, we propose Self-Adaptive Domain Adaptation (SADA) from black-box predictors based on the model’s learning status, tracking the learning status of each class, introducing class-specific thresholds to filter out noisy labels, and employ an adaptive target domain data distribution estimation to construct a balanced class distribution while avoiding data scarcity. Experimental results demonstrate that our method not only mitigates confirmation bias but also helps in model training. We achieve state-of-the-art results on the Office-31, Office-Home, and VisDA-C datasets.