A novel deep transfer learning method based on explainable feature extraction and domain reconstruction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep transfer learning faces challenges with "black-box" models and unstable feature adaptation.
• Explainability and reliable feature adaptation are crucial for advancing transfer learning.

Purpose of the Study:

• To propose a multi-stage deep transfer learning method (XDTL) that enhances model performance and explainability.
• To address the limitations of current deep transfer learning techniques through explainable feature extraction and domain reconstruction.

Main Methods:

• XDTL divides features into key and regular types using cross-validation and explainability analysis.
• It employs a seed replacement strategy with key target samples for target domain reconstruction.
• This approach facilitates a deep transfer process.

Main Results:

• XDTL demonstrated an average effectiveness improvement of 27.43% compared to existing methods.
• The proposed method shows superior performance and enhanced explainability in transfer learning tasks.

Conclusions:

• XDTL offers a promising solution for the explainability challenges in deep transfer learning.
• The method provides new insights and potential for diverse applications across various machine learning tasks.