Semi-Supervised Domain Adaptive Structure Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised domain adaptation (SSDA) faces challenges with overfitting to limited labeled data and distribution shifts between domains.
• Existing methods combining domain adaptation (DA) and semi-supervised learning (SSL) often struggle due to training data bias towards labeled samples.

Purpose of the Study:

• To introduce an adaptive structure learning method that regularizes the interplay between SSL and DA for improved SSDA.
• To enhance representation learning by increasing intra-class density and decision boundary smoothness in cross-domain scenarios.

Main Methods:

• A framework utilizing a shared feature encoder and two classifiers with contradictory objectives: one for target feature clustering and another for source feature scattering.
• Integration of Maximum Mean Discrepancy (MMD) for cross-domain feature alignment and self-training (ST) for leveraging partially labeled data.
• Multi-view learning principles to reconcile conflicting objectives within a shared feature space.

Main Results:

• Demonstrated accuracy and robustness on standard SSDA benchmarks like DomainNet and Office-home.
• The proposed method effectively addresses overfitting and distribution shift issues inherent in SSDA.
• Achieved superior performance compared to existing state-of-the-art approaches in SSDA.

Conclusions:

• The adaptive structure learning method provides a robust solution for semi-supervised domain adaptation.
• The framework's ability to regularize SSL and DA cooperation leads to more reliable cross-domain predictions.
• This work advances SSDA by effectively aligning features and learning from partially labeled data.