Transferring Knowledge Fragments for Learning Distance Metric from a Heterogeneous Domain

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning enhances target tasks using knowledge from related tasks.
• Distance Metric Learning (DML) often faces label deficiency issues.
• Existing Transfer DML (TDML) methods struggle with heterogeneous domains.

Purpose of the Study:

• To develop a general and flexible heterogeneous transfer DML (HTDML) framework.
• To overcome limitations of current TDML and heterogeneous transfer learning (HTL) methods.
• To enable effective knowledge transfer across heterogeneous domains for DML.

Main Methods:

• A novel HTDML framework is proposed.
• Any DML algorithm can pre-learn the source metric.
• Pre-learned source metrics are used as knowledge fragments for target learning.
• The framework supports both linear and nonlinear target metric learning.

Main Results:

• The proposed transfer strategy reduces generalization error in the target domain.
• Novel algorithms are developed for learning linear or nonlinear target metrics.
• Extensive experiments demonstrate the method's effectiveness across various applications.

Conclusions:

• The developed HTDML framework offers a flexible and effective solution for cross-domain DML.
• The method successfully transfers knowledge from source to target domains, even with heterogeneous data.
• This approach advances the field of transfer learning for distance metric learning.