Graphical Modeling for Multi-Source Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Multi-Source Domain Adaptation (MSDA) addresses knowledge transfer from multiple sources to a target domain, a more practical yet challenging problem than single-source adaptation.
• Effective joint modeling of multiple source and target domains, along with domain combination, is crucial for MSDA.
• Graphical structures can effectively model interdependencies among instances and categories across domains.

Purpose of the Study:

• To propose novel graphical models for MSDA that facilitate cross-domain joint modeling and learnable domain combination.
• To introduce Conditional Random Field for MSDA (CRF-MSDA) and Markov Random Field for MSDA (MRF-MSDA).

Main Methods:

• CRF-MSDA learns the joint distribution of labels conditioned on observations by constructing a relational graph and using local message passing.
• MRF-MSDA models the joint distribution of observations over Markov networks using an energy-based formulation, enabling label prediction via likelihood summation.
• MRF-MSDA offers greater expressiveness and lower computational cost compared to CRF-MSDA.

Main Results:

• Both CRF-MSDA and MRF-MSDA achieved superior performance over existing methods on four standard MSDA benchmark datasets.
• The models demonstrated effectiveness across datasets with varying domain shifts and data complexities.
• Analytical studies provided insights into the performance of cross-domain joint modeling and the impact of different model components.

Conclusions:

• The proposed graphical models, CRF-MSDA and MRF-MSDA, offer effective solutions for Multi-Source Domain Adaptation.
• MRF-MSDA presents a more expressive and computationally efficient approach for MSDA.
• The findings highlight the benefits of graphical structures for joint domain modeling and knowledge transfer in MSDA.