A Bayesian approach to semi-supervised domain adaptation in streaming data

One of the main challenges in statistical learning with large datasets is heterogeneity, which can violate the conventional assumption that training and test data are drawn from the same distribution. To address these distribution shifts, domain adaptation techniques have been developed to accurately learn a concept representing the source domain and transfer this knowledge to the target domain, where labels might not be fully available. This work addresses the challenge of adapting to domain shifts in streaming data for classification tasks, where labels in the target domain are only partially observed. It treats these distribution shifts as concept drift, referring to the changes in the statistical properties of the data over time. We propose an ensemble-based approach within the Bayesian regularization framework. The proposed approach adapts to various drifts by adjusting the weights of the combiner to reflect the relevance of the predicted models. It also provides an uncertainty measure for the final prediction, enhancing the reliability. Leveraging this uncertainty, we treat unlabelled instances in the target domain with high-confidence predictions as pseudo-labelled data and update the ensemble model accordingly. The effectiveness of our proposed method is demonstrated through numerical examples.