A Hybrid Evolutionary Computation Approach to Inducing Transfer Classifiers for Domain Adaptation

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Domain adaptation aims to leverage knowledge from a source domain to improve performance in a related target domain.
• Traditional methods often align domains before classification, while transfer classifier induction integrates alignment into classifier construction.
• Existing transfer classifier induction methods, primarily gradient-based, risk getting stuck in local optima.

Purpose of the Study:

• To propose a novel transfer classifier induction algorithm using evolutionary computation to overcome limitations of gradient-based approaches.
• To introduce a low-dimensionality representation for transfer classifiers.
• To develop a hybrid optimization process balancing manifold consistency and domain difference.

Main Methods:

• A transfer classifier induction algorithm based on evolutionary computation.
• A novel, lower-dimensional representation for transfer classifiers.
• A hybrid optimization process combining evolutionary search for manifold consistency and gradient-based local search for domain difference reduction.

Main Results:

• The proposed algorithm achieved superior performance compared to seven traditional and four deep domain adaptation methods.
• The evolutionary approach effectively addressed the local optima issue inherent in gradient-based methods.
• The novel representation and hybrid optimization successfully balanced key domain adaptation objectives.

Conclusions:

• Evolutionary computation offers a robust alternative for transfer classifier induction in domain adaptation.
• The proposed method demonstrates significant improvements in classification performance across domains.
• This work advances domain adaptation techniques by integrating global search with local refinement.