Deep Decision Tree Transfer Boosting.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Instance transfer learning methods augment target data with source data when target labels are scarce.
• Boosting-based transfer learning, like TrAdaBoost, is common but can overfit complex data.
• Deep decision trees offer complex hypotheses but risk overfitting in transfer learning.

Purpose of the Study:

• To propose a new instance transfer learning method, Deep Decision Tree Transfer Boosting (DTrBoost).
• To address overfitting issues in boosting-based transfer learning with deep decision trees.
• To enable learning deep decision trees without overfitting in transfer learning scenarios.

Main Methods:

• Developed DTrBoost, a novel instance transfer learning algorithm.
• Learned base learner weights by minimizing data-dependent learning bounds across source and target domains.
• Utilized Rademacher complexities to ensure generalization and prevent overfitting.

Main Results:

• Demonstrated that DTrBoost can learn deep decision trees without overfitting.
• Theorem proofs and experimental results confirm the method's effectiveness.
• Successfully applied instance transfer learning to complex data with limited labels.

Conclusions:

• DTrBoost offers an effective solution for transfer learning with deep decision trees.
• The method provides a robust approach to mitigate overfitting in complex datasets.
• This work advances transfer learning techniques for improved model generalization.