Instance transfer learning with multisource dynamic TrAdaBoost.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Transfer learning offers advantages over traditional methods by reducing costs and improving efficiency.
• Challenges arise when source and target domain data distributions are similar, necessitating advanced transfer techniques.

Purpose of the Study:

• To propose a novel instance transfer learning method utilizing multisource dynamic TrAdaBoost.
• To address the challenge of similar data distributions between source and target domains.
• To enhance classification accuracy and learning efficiency in target tasks.

Main Methods:

• Developed a multisource dynamic TrAdaBoost algorithm for instance transfer learning.
• Incorporated a dynamic factor to mitigate weight entropy drift between source and target instances.
• Utilized knowledge from multiple source domains to prevent negative transfer and train candidate classifiers.

Main Results:

• Theoretical analysis indicates improved capability in managing weight entropy drift.
• Experimental results demonstrate superior classification effectiveness compared to single-source transfer learning.
• The proposed algorithm achieved higher classification accuracy on target tasks.

Conclusions:

• The multisource dynamic TrAdaBoost method effectively transfers knowledge from multiple domains.
• The dynamic factor enhances the algorithm's robustness and performance.
• This approach offers a significant improvement in classification accuracy for transfer learning scenarios.