Genetic Programming for Document Classification: A Transductive Transfer Learning System.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Document classification faces challenges due to high-dimensional and sparse data.
• Transfer learning aims to improve classification by transferring knowledge between domains.
• Existing methods often struggle with unlabeled target domain data.

Purpose of the Study:

• To propose a novel transductive transfer learning system for document classification.
• To address the challenge of unlabeled training data in the target domain.
• To leverage genetic programming (GP) for automatic pseudolabeling.

Main Methods:

• A transductive transfer learning system utilizing genetic programming (GP) solutions.
• Automatic pseudolabeling of target domain training data.
• Retraining classifiers using all target domain features.

Main Results:

• The proposed transductive GP system demonstrated superior prediction accuracy on target domain test data.
• Outperformed existing transfer learning approaches, including subspace alignment and feature-level domain adaptation methods.
• Showed significant improvements over a recent pseudolabeling strategy-based method.

Conclusions:

• The developed transductive GP system effectively handles document classification with unlabeled target data.
• This approach offers a robust solution for improving transfer learning performance.
• The method shows promise for real-world applications requiring accurate classification with limited labeled data.