A hybrid framework for unsupervised domain adaptation: Integrating feature alignment, parameter adaptation, and instance selection

Abstract Unsupervised Domain Adaptation (UDA), a key subfield of Transfer Learning (TL), addresses the performance degradation caused by domain shift that occurs when a model trained on a labeled source domain is applied to an unlabeled target domain. Despite considerable advances, most existing UDA methods suffer from two major limitations: (1) heavy reliance on extracting domain-invariant features without explicitly enhancing their discriminative power for classification or recognition tasks, and (2) confinement to a single adaptation strategy-such as feature-based, instance-based, or parameter-based-overlooking the potential advantages of combining multiple adaptation paradigms. To tackle these issues simultaneously, this paper proposes a novel Hybrid Adaptation of Feature, Parameter, and Instance (HA-FPI) framework for UDA. HA-FPI integrates feature alignment, parameter adaptation, and instance selection into a unified optimization objective, facilitating the extraction of cross-domain invariant yet discriminative features, optimizing shared parameters, and selecting high-confidence target samples. This integrated approach mitigates feature distribution discrepancy and significantly enhances cross-domain knowledge transfer efficiency. The proposed method comprises three stages: discriminative cross-domain invariant feature extraction, classifier parameter adaptation, and fused decision-making. In the first stage, the Cross-Domain Mean Approximation (CDMA) distance is minimized to identify a low-dimensional feature subspace that reduces both marginal and conditional distribution discrepancy between domains. Additionally, CDMA is combined with linear interpolation to construct a cross-domain linear sample generation mechanism, producing transitional samples that smooth the knowledge transfer process. To improve the discriminability of the shared feature representations, a discriminative penalty constraint is applied to both original and synthesized samples, effectively separating heterogeneous samples in the feature subspace. During the classifier parameter adaptation stage, a Least Squares Regression (LSR) is trained on source samples within the shared feature subspace. A unidirectional CDMA is incorporated into its objective function, yielding a source LSR with inherent transfer capability. Using CDMA as a confidence measure, high-confidence target samples are selected and combined with source samples to train a target LSR. In the final fused decision stage, predictions for the test samples are obtained through the joint decisions of both LSRs. Extensive experiments on benchmark domain adaptation datasets including Office+Caltech256, Office-31, USPS+MNIST, and ImageCLEF demonstrate the superior performance of the proposed HA-FPI framework compared to state-of-the-art UDA methods.