Triplex transfer learning: exploiting both shared and distinct concepts for text classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Transfer learning addresses data distribution differences between source and target domains.
• Existing methods often assume shared concepts across domains, limiting transferability.
• High-level concepts like word clusters can model distribution differences but may miss domain-specific nuances.

Purpose of the Study:

• To propose a general transfer learning framework capable of exploring both shared and distinct concepts simultaneously.
• To improve classification accuracy by accommodating domain-specific features.
• To enhance prediction performance through manifold regularization.

Main Methods:

• A novel transfer learning framework based on nonnegative matrix trifactorization (NMTF).
• Simultaneous exploration of shared and distinct concepts across multiple data domains.
• Manifold regularization applied to target domains to improve generalization.
• Development and theoretical convergence analysis of an iterative optimization algorithm.

Main Results:

• The proposed NMTF-based framework significantly outperforms baseline transfer learning methods.
• The model demonstrates superior performance on challenging tasks with distinct domain concepts.
• Incorporating both shared and distinct concepts leads to greater flexibility and improved classification accuracy.

Conclusions:

• The developed transfer learning framework effectively models shared and distinct concepts for improved cross-domain knowledge transfer.
• The method offers enhanced flexibility and robustness, particularly for complex datasets with domain-specific characteristics.
• Future work could explore extensions of this framework to other machine learning tasks beyond classification.