Exploring Unsupervised Domain Adaptation Through the Lens of Unconstrained Label-Space Overlaps

In the rapidly evolving landscape of supervised learning, a significant reliance on large, annotated datasets has been both an asset and a challenge. Domain adaptation, particularly when delving into diverse label spaces, is a strategic countermeasure that enables models to transcend the limitations of the traditional closed-set paradigm. To address these challenges, this dissertation proposes three frameworks within the context of Partial Domain Adaptation that effectively mitigate negative transfer. A fourth method focusing on Universal Domain Adaptation is introduced, further expanding the scope and applicability of domain adaptation strategies presented in this work. The first framework features a dual-encoder setup with private and shared encoders designed to effectively capture domain-specific and shared features. This configuration employs a combination of soft and binary weights to prioritize relevant features, minimizing negative transfer, and integrates a complement entropy objective to reduce classification uncertainty. Advancing further, the second framework employs a category-level feature alignment technique that moves beyond first-order moments to delineate distinct category distributions. This method incorporates robust pseudo-labeling with an adaptive threshold for enhanced target supervision, validated by extensive benchmark testing against state-of-the-art models. The third approach is centered on a robust target-supervision strategy that utilizes ensemble learning on negative classes to refine pseudo-labels. This method ensures data privacy by eliminating the need for source data during the adaptation phase, making it particularly suited for sensitive applications such as in the medical field. A novel UDA method is also introduced, addressing overlapping and non-overlapping label spaces by employing ensemble learning and negative learning to generate robust pseudo-labels. This approach uses one-vs-all classifiers to identify unknown instances and employs an affinity metric for semantic similarity, ensuring accurate mapping and grouping of categories. The results of these efforts are underscored by thorough assessments of benchmark datasets, affirming the resilience, superior generalizability, and distinct advantage of the proposed frameworks over prevailing approaches. The continuation of this research aims to lay a robust groundwork for future explorations centered on domain adaptation across unconstrained label spaces, aiming to broaden the impact of the current discoveries.