Multisource Heterogeneous Domain Adaptation With Conditional Weighting Adversarial Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Heterogeneous domain adaptation (HDA) addresses challenges in cross-domain learning with differing data distributions and feature representations.
• Existing HDA research predominantly focuses on single-source scenarios, limiting applicability in real-world multisource data situations.

Purpose of the Study:

• To address the multisource heterogeneous domain adaptation problem.
• To propose a novel Conditional Weighting Adversarial Network (CWAN) for improved cross-domain learning.

Main Methods:

• CWAN employs adversarial learning to develop a feature transformer, label classifier, and domain discriminator.
• A sophisticated conditional weighting scheme quantifies source domain importance based on conditional distribution divergence between source and target domains.
• The proposed weighting scheme implicitly aligns conditional distributions during optimization, enhancing domain adaptation.

Main Results:

• CWAN demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on four real-world datasets, validating the effectiveness of the proposed approach.

Conclusions:

• The proposed CWAN effectively handles multisource heterogeneous domain adaptation.
• The conditional weighting scheme offers a significant advancement in aligning distributions and weighting source domains for improved adaptation performance.