Active learning for semi-supervised clustering based on locally linear propagation reconstruction.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Semi-supervised clustering performance depends heavily on the quality of provided side information.
• Existing methods may not optimally leverage pairwise constraints for effective clustering.

Purpose of the Study:

• To propose a novel active learner for generating effective pairwise constraints (must-link and cannot-link) to enhance semi-supervised clustering.
• To develop techniques for identifying less informative samples and optimizing the selection of new constraints.

Main Methods:

• Utilized a kernel version of locally linear embedding for manifold learning to identify unimportant samples.
• Developed a query selection criterion considering sample importance and expected query cost.
• Introduced a technique to infer must-links from manifold information for semi-supervised clustering algorithms.

Main Results:

• The proposed methods effectively identify samples that do not significantly contribute to cluster structures.
• The novel query selection strategy balances information gain and efficiency.
• Inferred must-links successfully convey manifold structure information to clustering algorithms.

Conclusions:

• The developed active learner generates effective pairwise constraints that capture underlying cluster structures.
• The proposed approach demonstrates feasibility and improves semi-supervised clustering performance.