Transfer Kernel Learning for Multi-Source Transfer Gaussian Process Regression.

Area of Science:

• Machine Learning
• Statistics
• Computer Science

Background:

• Multi-source transfer regression presents challenges in adaptive knowledge transfer due to the need to model diverse domain relatedness.
• Existing transfer kernels have limitations in effectively handling the complexities of multi-source transfer regression problems.

Purpose of the Study:

• To propose an effective method for explicitly modeling domain relatedness in multi-source transfer regression using transfer kernel learning.
• To introduce a novel multi-source transfer kernel (k_ms) that addresses limitations of existing approaches.

Main Methods:

• The proposed k_ms kernel assigns learnable coefficients to model inter-domain relatedness while enforcing intra-domain relatedness to 1.
• k_ms utilizes different standard kernels for different domain pairs to capture heterogeneous data characteristics.
• A theorem is provided to guarantee the positive semi-definiteness of k_ms and offer semantic interpretation of learned domain relatedness.

Main Results:

• The proposed k_ms kernel effectively models domain relatedness by assigning learnable parametric coefficients.
• The method successfully captures heterogeneous data characteristics across multiple domains.
• Empirical studies demonstrate the effectiveness of the proposed method in improving transfer performance.

Conclusions:

• The novel multi-source transfer kernel (k_ms) offers an effective solution for modeling domain relatedness in multi-source transfer regression.
• The proposed method enhances adaptive knowledge transfer by accurately capturing inter-domain relationships.
• The theoretical guarantees and empirical results validate the utility of k_ms in transfer learning applications.