L1-norm locally linear representation regularization multi-source adaptation learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Supervised domain adaptation learning (DAL) often faces challenges with limited labeled target domain data.
• Effective utilization of abundant unlabeled data is crucial for generalization in small-sample regimes.

Purpose of the Study:

• To develop a novel framework for supervised DAL that leverages geometric properties of the target marginal distribution.
• To enhance generalization by incorporating target geometric structure into existing model-based DAL methods.

Main Methods:

• Proposed a novel L1-norm locally linear representation (L1-LLR) regularization framework for robust graph construction.
• Introduced a new L1-LLR graph Laplacian regularization for semi-supervised learning with multi-source adaptation constraints (L1-MSAL).
• Generalized L1-MSAL to handle nonlinear learning problems using a reproducing kernel Hilbert space (RKHS) mapping.

Main Results:

• The L1-MSAL method demonstrated promising experimental results on real-world datasets.
• The framework effectively exploits the geometry of probability distributions for improved adaptation.
• Achieved successful adaptation in scenarios with limited labeled target data.

Conclusions:

• The proposed L1-MSAL framework offers an effective approach for supervised domain adaptation, particularly in small-sample settings.
• Incorporating geometric information of the target marginal distribution enhances model generalization.
• The method shows potential for various applications including face recognition and object recognition.