Pairwise Constraint Propagation With Dual Adversarial Manifold Regularization.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Pairwise constraints (PCs), including must-links (MLs) and cannot-links (CLs), are crucial for semisupervised learning.
• Existing pairwise constraint propagation (PCP) methods often use a single matrix, failing to differentiate between MLs and CLs, thus limiting discriminability.

Purpose of the Study:

• To propose a novel PCP model using dual adversarial manifold regularization to effectively utilize limited initial PCs.
• To enhance the discriminability of propagated PCs by separately handling MLs and CLs.

Main Methods:

• Developed a novel PCP model employing dual adversarial manifold regularization.
• Propagated MLs and CLs using separate similarity and dissimilarity matrices guided by a data sample graph structure.
• Incorporated an adversarial relationship between the similarity and dissimilarity matrices.
• Formulated the model as a nonnegative constrained minimization problem solvable with guaranteed convergence.

Main Results:

• The proposed model demonstrates superior performance compared to state-of-the-art PCP models in extensive experiments.
• Validated effectiveness in PC propagation and applications like constrained clustering and metric learning.

Conclusions:

• The dual adversarial manifold regularization approach effectively leverages limited pairwise constraints.
• The proposed method offers improved discriminability and performance in semisupervised tasks.