Practical Multisource Transfer Regression With Source-Target Similarity Captures.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Multisource transfer learning faces challenges in accurately modeling diverse source-target similarities.
• Existing Gaussian Process (GP) models often struggle with scalability and capturing unique similarities across multiple sources.

Purpose of the Study:

• To develop an advanced framework for multisource transfer regression that explicitly models varied source-target similarities.
• To address the limitations of current GP-based approaches in handling complex transfer learning scenarios.

Main Methods:

• Investigated transfer covariance functions for GP models to represent source-target domain similarities.
• Proposed TCMSStack, an integrated framework using separate transfer covariance functions for each source with a stacking approach.
• Developed two instances of the TCMSStack framework.

Main Results:

• Demonstrated that general GP models with a single transfer covariance function yield suboptimal performance due to uniform similarity coefficients.
• TCMSStack effectively learns source-target similarities by considering inter-source dependencies within base GP models.
• Experimental results on synthetic and real-world datasets confirmed the effectiveness of the proposed approach, even with up to 11 sources.

Conclusions:

• TCMSStack offers a significant advancement in multisource transfer regression by enabling nuanced modeling of source-target relationships.
• The framework overcomes scalability issues and improves transfer performance compared to traditional methods.
• The proposed approach provides a robust solution for complex transfer learning tasks with multiple diverse data sources.