Deep Multi-Source Supervised Domain Adaptation with Class Imbalance

AbstractDeep multi-source domain adaptation is based on deep neural networks and exploits knowledge from multiple source domains to improve predictions on a target domain.In this paper, we are especially interested in investigating domain adaptation in a supervised context, with limited data and class imbalance. We propose a new multi-source supervised domain adaptation approach that is able to transfer both shared knowledge across all source domains and source domain specific knowledge toward a target domain. Transfer contribution weights are computed during training based on domain divergence. They are used to balance each source domain impact on learning during the model training phase, limiting negative transfer as much as possible. We conduct extensive experiments to show that our approach competes and even outperforms other state-of-the-art domain adaptation approaches on both image benchmark datasets and real-world tabular medical data. We perform statistical analysis to better evaluate our experimental results, and conduct an ablation study to evaluate the usefulness of each component of the method.