Regularized Multi-Output Gaussian Convolution Process With Domain Adaptation.

Area of Science:

• Machine Learning
• Statistical Modeling
• Transfer Learning

Background:

• Multi-output Gaussian Process (MGP) is a powerful transfer learning technique for modeling multiple related outputs.
• Existing MGP methods face challenges with negative transfer (lack of shared information) and input domain inconsistency.
• These challenges limit the effectiveness of MGP in diverse real-world applications.

Purpose of the Study:

• To propose a regularized MGP framework with domain adaptation to overcome negative transfer and input domain inconsistency.
• To enhance knowledge transfer by adaptively selecting informative outputs and aligning input domains across different tasks.

Main Methods:

• Developed a sparse MGP covariance matrix using a convolution process with penalization terms for adaptive output selection.
• Implemented a domain adaptation method by marginalizing inconsistent features and expanding missing features to align input domains.
• Provided theoretical guarantees on the statistical properties for practical and asymptotic performance.

Main Results:

• The proposed framework significantly outperformed state-of-the-art benchmarks in extensive simulation studies.
• Demonstrated effectiveness in a real-world case study involving a ceramic manufacturing process.
• Successfully addressed both negative transfer and input domain inconsistency issues.

Conclusions:

• The novel MGP framework effectively handles negative transfer by adaptively identifying informative outputs.
• The integrated domain adaptation successfully aligns input domains, improving transfer learning performance.
• The proposed method offers a robust solution for complex transfer learning scenarios with heterogeneous data.