Multi-EPL: Accurate multi-source domain adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Multi-source domain adaptation (MSDA) addresses training models with labeled source data for unlabeled target data.
• Existing MSDA methods often fail to fully leverage target data or align features effectively.
• Privacy concerns frequently limit the availability of labeled target data in practical applications.

Purpose of the Study:

• To develop a novel method, Multi-EPL, for effective multi-source domain adaptation (MSDA) when target data lacks labels.
• To overcome limitations of existing MSDA frameworks by incorporating label-specific feature alignment and ensemble feature extraction.
• To improve the performance of models in scenarios with heterogeneous source datasets and unlabeled target datasets.

Main Methods:

• Multi-EPL employs label-wise moment matching to align conditional feature distributions across domains.
• The method utilizes pseudo-labeling techniques to leverage unlabeled target data.
• An ensemble of multiple feature extractors is integrated to enhance domain adaptation robustness and accuracy.

Main Results:

• Multi-EPL achieved state-of-the-art performance on MSDA tasks across both image and text domains.
• The proposed method demonstrated significant accuracy improvements, reaching up to 13.20%.
• Experiments validated the effectiveness of label-wise alignment, pseudo-labeling, and feature extractor ensembles.

Conclusions:

• Multi-EPL offers a superior approach to multi-source domain adaptation with unlabeled target data.
• The method's components effectively address key challenges in domain adaptation, improving model generalization.
• This research provides a valuable contribution to machine learning, particularly in privacy-preserving and data-scarce scenarios.