Simultaneous Dimensionality Reduction and Classification via Dual Embedding Regularized Nonnegative Matrix...

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Nonnegative Matrix Factorization (NMF) is a common technique for data representation.
• Traditional NMF-based classification uses a stepwise approach, separating NMF from classification.
• This separation can lead to suboptimal classification accuracy due to overlooked dependencies.

Purpose of the Study:

• To unify NMF and classification processes for improved accuracy.
• To introduce a novel semi-supervised method, Dual Embedding Regularized NMF (DENMF).
• To simultaneously optimize low-dimensional representations and class assignments.

Main Methods:

• Formulated a novel constrained optimization model: Dual Embedding Regularized NMF (DENMF).
• Employed joint optimization to simultaneously learn low-dimensional representations and an assignment matrix.
• Utilized locally linear embedding to preserve geometric structure across feature and label spaces.
• Developed an alternating iteration algorithm with theoretically proven convergence.

Main Results:

• DENMF effectively unifies feature learning and classification within a single framework.
• The method achieved higher classification accuracy compared to state-of-the-art algorithms.
• Experimental validation was performed on five benchmark datasets.

Conclusions:

• The proposed DENMF method offers a more effective approach to semi-supervised classification.
• Jointly optimizing representations and assignments leads to significant improvements in accuracy.
• DENMF demonstrates superior performance over existing NMF-based and other classification methods.