Partial Domain Adaptation via Importance Sampling-Based Shift Correction.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Partial Domain Adaptation (PDA) addresses knowledge transfer from labeled source to unlabeled target domains.
• Existing PDA methods often use sample reweighing, which can lead to overfitting and underutilization of labeled data.
• A key challenge in PDA is correcting label distribution shifts while preserving model generalization.

Purpose of the Study:

• To propose a novel Importance Sampling-based Shift Correction (IS2C) method for Partial Domain Adaptation.
• To enhance the generalization ability of machine learning models in PDA scenarios.
• To provide theoretical guarantees and practical improvements over existing PDA techniques.

Main Methods:

• Developed IS2C, which samples new labeled data from a constructed sampling domain with a target-like distribution.
• Incorporated mixture distribution sampling to link domain shift to generalization error, offering interpretability.
• Utilized an optimal transport-based independence criterion for conditional distribution alignment, with complexity optimized to O(n^2).

Main Results:

• IS2C demonstrates theoretical guarantees, proving that generalization error can be effectively controlled.
• Experiments on PDA benchmarks validate the theoretical findings.
• The proposed IS2C method shows superior performance compared to existing PDA techniques.

Conclusions:

• IS2C offers a robust approach to Partial Domain Adaptation by addressing label distribution shifts.
• The method enhances model generalization and reduces overfitting through strategic data sampling.
• IS2C represents a significant advancement in knowledge transfer for machine learning applications with domain shifts.