Learning Transferred Weights From Co-Occurrence Data for Heterogeneous Transfer Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Heterogeneous transfer learning faces challenges in determining effective knowledge transfer between domains.
• Quantifying the degree of knowledge transfer is crucial for optimal model performance.

Purpose of the Study:

• To propose a novel method for evaluating domain relatedness in heterogeneous transfer learning.
• To learn transferred weights that quantify the effectiveness and amount of knowledge transfer.

Main Methods:

• Utilizing co-occurrence data with instances in different feature spaces.
• Computing principal components to represent co-occurrence data across feature spaces.
• Employing Markov Chain Monte Carlo to construct a directed cyclic network representing domain dependencies.
• Using edge weights as transferred weights for knowledge transfer control.

Main Results:

• The proposed method effectively captures varying degrees of relations among feature spaces.
• Experimental results demonstrate enhanced learning performance in heterogeneous transfer learning.
• Learned weights serve as priors to control knowledge transfer in existing methods.

Conclusions:

• The novel approach accurately assesses domain relatedness through learned transferred weights.
• This method provides a robust mechanism for optimizing knowledge transfer in heterogeneous settings.
• The findings contribute to advancing the capabilities of heterogeneous transfer learning models.