Max-min distance nonnegative matrix factorization.

Area of Science:

• Machine Learning
• Data Science
• Pattern Recognition

Background:

• Nonnegative Matrix Factorization (NMF) is a common technique for data representation in pattern classification.
• Traditional NMF methods do not utilize class label information.
• This limitation hinders the discriminative power of NMF-derived representations.

Purpose of the Study:

• To develop a novel supervised Nonnegative Matrix Factorization (NMF) algorithm that incorporates class labels.
• To enhance the discriminative ability of NMF representations for pattern classification tasks.
• To improve the performance of NMF by explicitly considering within-class and between-class relationships.

Main Methods:

• The proposed algorithm leverages class labels to categorize data sample pairs into within-class and between-class sets.
• An objective function is constructed to minimize the maximum distance within-class pairs and maximize the minimum distance between-class pairs in the NMF space.
• An iterative optimization process is employed, updating basis matrices, coefficient matrices, and slack variables.

Main Results:

• The supervised NMF algorithm demonstrates improved discriminative ability in the new representation space.
• Experimental evaluations on three pattern classification problems show superior performance compared to state-of-the-art supervised NMF methods.
• The proposed method effectively utilizes class label information to refine data representations.

Conclusions:

• The novel supervised NMF algorithm effectively enhances pattern classification by incorporating class label information.
• The method's objective function, focusing on within-class compactness and between-class separability, leads to more discriminative representations.
• This approach represents a significant advancement in supervised dimensionality reduction and feature learning for classification.