Domain Adaption via Feature Selection on Explicit Feature Map.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation aims to improve model performance on a target domain using data from a source domain.
• Current methods often use all features, which can degrade performance if irrelevant features are included.
• Invariant features are crucial for effective domain adaptation but may exist in higher-order feature spaces.

Purpose of the Study:

• To propose a novel domain adaptation algorithm that effectively selects invariant and discriminative features.
• To address the limitation of using all features in traditional domain adaptation approaches.
• To reveal high-order invariant features and filter weak ones for improved cross-domain generalization.

Main Methods:

• Representing data using a kernel-induced explicit feature map to reveal high-order invariant features.
• Minimizing marginal and conditional distribution differences along with model error for feature selection.
• Employing a cutting plane algorithm to solve the NP-hard feature selection problem.

Main Results:

• The proposed algorithm effectively selects invariant discriminative features.
• Experimental results on six real-world benchmarks demonstrate superior performance compared to state-of-the-art methods.
• The algorithm shows both effectiveness and efficiency in domain adaptation tasks.

Conclusions:

• The novel domain adaptation algorithm successfully identifies beneficial invariant features.
• This approach enhances cross-domain generalization by avoiding the negative impact of irrelevant features.
• The method offers a significant improvement over existing domain adaptation techniques.