Semi-supervised non-negative matrix factorization with structure preserving for image clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Semi-supervised learning effectively uses partial data labels.
• Non-negative matrix factorization (NMF) is valued for its interpretability and practicality.
• Existing semi-supervised NMF methods often overlook NMF's inherent structure.

Purpose of the Study:

• To propose a novel semi-supervised NMF method that preserves the intrinsic structure of NMF.
• To improve the utilization of label information while maintaining NMF's structure.
• To enhance the performance of image clustering tasks.

Main Methods:

• A new weighted label matrix and label constraint regularizer are constructed.
• Basis images of labeled data are extracted to guide the learning of all basis images.
• A basis regularizer is established to monitor and modify basis image learning.
• The proposed method incorporates both label constraint and basis regularizers into NMF.
• A multiplicative updating algorithm is developed for optimization.

Main Results:

• The proposed semi-supervised NMF method demonstrates effectiveness in image clustering.
• Experimental results on eight datasets show superior performance compared to existing algorithms.
• The method successfully balances label utilization with the preservation of NMF's intrinsic structure.

Conclusions:

• The novel semi-supervised NMF method offers significant improvements for image clustering.
• Preserving the intrinsic structure of NMF is crucial for effective semi-supervised learning.
• The approach provides a valuable advancement in semi-supervised learning techniques.