DANE: A Dual-Level Alignment Network With Ensemble Learning for Multisource Domain Adaptation

Multisource domain adaptation (MDA) aims to transfer knowledge from multiple labeled source domains to an unlabeled target domain. However, the severe intradomain and interdomain shifts exacerbate the difficulty of knowledge transfer. Besides, the potential knowledge embedded within and across source domains is often ignored by past work. In this article, we propose a dual-level alignment network with ensemble learning (DANE) to address the above limitations for MDA classification tasks. Concretely, to deal with the intradomain shift, we present an intradomain alignment by employing a novel clustering loss with a batchwise prediction similarity matrix (PSM) to foster a discriminative and well-clustered target feature space inside the target domain. To tackle the interdomain shift, we introduce an interdomain alignment loss with a dynamic weighting function to allocate greater weights to samples with higher uncertainty, thus leading to a more precise alignment between the source and target distributions. Moreover, to comprehensively harness the inherent knowledge present within and across the source domains, we leverage the principles of ensemble learning. This involves constructing an ensemble source that amalgamates all individual source domains, enabling us to fully explore the domain-invariant knowledge embedded within the source data. Subsequently, we employ an ensemble-weighted decision integration strategy, which facilitates the fusion of beneficial knowledge from each source domain, resulting in an optimal exploration and utilization of the domain-invariant knowledge for enhanced classification performance. Extensive experiments on five MDA benchmark datasets demonstrate that DANE outperforms other methods by a significant margin, defining the new state of the art (SOTA). Moreover, we also validate the proposed DANE in open-world scenarios, which further shows the effectiveness of our method.