A New Belief-Based Bidirectional Transfer Classification Method

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Transfer learning addresses classification with limited target domain data by leveraging related source domain data.
• Heterogeneous features and uncertainty in domain transformation can hinder classification accuracy.
• Effective uncertainty management is crucial for improving classification performance in transfer learning scenarios.

Purpose of the Study:

• To propose a novel belief-based bidirectional transfer classification (BDTC) method to enhance classification accuracy.
• To effectively manage uncertainty inherent in domain transformation during transfer learning.
• To leverage complementary knowledge from both source and target domains for improved classification.

Main Methods:

• BDTC estimates an intraclass transformation matrix to map source domain patterns to the target domain.
• It transfers labeled source domain patterns to the target domain for classifier training.
• Query patterns are transferred from target to source domains using K-NN, and results are combined using belief functions theory for weighted combination.

Main Results:

• The proposed BDTC method achieves improved classification accuracy compared to existing methods.
• It effectively reduces uncertainty in transfer classification through a novel combination strategy.
• Experiments on domain adaptation benchmarks validate the method's effectiveness.

Conclusions:

• The belief-based bidirectional transfer classification (BDTC) method offers a robust approach to handle heterogeneous features and uncertainty in transfer learning.
• Combining classification results from bidirectional transfers using belief functions theory significantly enhances accuracy.
• BDTC provides an effective strategy for domain adaptation tasks where labeled data is scarce in the target domain.