Probability-Based Graph Embedding Cross-Domain and Class Discriminative Feature Learning for Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Feature-based domain adaptation aims to align data distributions across domains for effective model transfer.
• Reducing domain shift and enhancing feature discriminability are critical challenges in unsupervised domain adaptation.

Purpose of the Study:

• To propose a unified framework, Probability-based Graph embedding Cross-domain and class Discriminative (PGCD), for unsupervised domain adaptation.
• To improve the alignment of local and global geometric structures across domains and enhance class-specific feature compactness.

Main Methods:

• Developed novel graph embedding structures for class-discriminative transfer learning and cross-domain alignment.
• Incorporated theoretical analyses to interpret the proposed graph structures and sample relationships.
• Utilized probability-based weight strategies to generate robust centroids, reducing error accumulation.

Main Results:

• The proposed PGCD framework demonstrated promising performance on benchmark datasets.
• The novel graph embedding structures effectively aligned geometric structures and improved feature compactness.
• The probability-based weighting strategy enhanced the accuracy of transfer feature learning.

Conclusions:

• The PGCD framework offers an effective solution for unsupervised domain adaptation by addressing domain shift and feature discriminability.
• The graph embedding approach provides interpretability for cross-domain transfer learning scenarios.
• Experimental results validate the superiority of PGCD over existing advanced methods.