Softly Associative Transfer Learning for Cross-Domain Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Cross-domain text classification faces challenges in applying models trained on one domain to another.
• Existing transfer learning methods often assume identical knowledge transfer matrices, which is unrealistic and limits performance.
• The presence of noisy labels in source domains further complicates accurate cross-domain classification.

Purpose of the Study:

• To propose a novel softly associative transfer learning algorithm for cross-domain text classification.
• To address the limitations of existing methods by allowing diversity in knowledge transfer and handling noisy labels.
• To develop an effective algorithm that improves classification performance across different domains.

Main Methods:

• Integration of two non-negative matrix tri-factorizations into a joint optimization framework.
• Inclusion of approximate constraints on word clusters and association matrices for diverse knowledge transfer.
• Incorporation of approximate constraints on source domain class labels to mitigate the impact of noisy labels.

Main Results:

• The proposed algorithm demonstrates effectiveness across various text datasets.
• Theoretical and empirical verification of the convergence of the iterative algorithm.
• Superior performance compared to state-of-the-art competitors in cross-domain text classification tasks.

Conclusions:

• The softly associative transfer learning algorithm offers a more realistic and effective approach to cross-domain text classification.
• The method's ability to handle label noise and allow for diversity in knowledge transfer enhances its robustness.
• This work provides a significant advancement in transfer learning for text classification applications.