Adversarial Domain Adaptation for Classification with Nested Dichotomies

Domain shift occurs when the data distribution used for testing diverges from the one employed during training, resulting in reduced performance of machine learning models. Domain adaptation (DA) offers techniques to mitigate disparities between two domains and enhance a model's effectiveness in the target domain. Most existing DA techniques have primarily focused on one-level classifiers (OLC), which are designed for non-hierarchical data. However, there is limited research on DA methods specifically tailored for hierarchical settings featuring a label hierarchy, such as Nested Dichotomies (NDCs) which transform a multiclass classification problem into a series of binary problems. In this paper, we introduce a discriminative domain matching approach tailored for NDCs, which can accommodate labels at various hierarchical levels. Our approach involves the development of an adversarial DA technique aimed at learning an invariant feature representation across the diverse hierarchical levels within the NDC framework. To achieve this, we partition the unlabeled target domain data at different levels of the hierarchy, guided by predicted labels obtained from previous levels. Our experiments on digit datasets demonstrate that our proposed algorithm outperforms the basic NDC model as it achieves higher classification accuracy when applied to data from the target domain.