Two-Dimensional Semi-Nonnegative Matrix Factorization for Clustering

Area of Science:

• Machine Learning
• Data Mining
• Dimensionality Reduction

Background:

• Existing 2D data factorization methods often lose crucial spatial information.
• Preprocessing 2D data into vectors can degrade its inherent structure.

Purpose of the Study:

• To propose a novel Semi-Nonnegative Matrix Factorization (TS-NMF) method for 2D data.
• To preserve spatial information lost in conventional vectorization approaches.
• To enhance data representation for improved clustering and analysis.

Main Methods:

• Developed a TS-NMF method integrating projection matrix seeking, new data representation building, and manifold learning.
• Constructed adaptive manifolds in projected subspaces to mitigate noise and outliers.
• Optimized projection directions guided by clustering objectives.

Main Results:

• TS-NMF effectively retains spatial information in 2D data representations.
• The integrated model yields powerful and representative data features.
• Experimental results demonstrate superior performance compared to state-of-the-art algorithms.

Conclusions:

• TS-NMF offers a significant advancement in analyzing 2D data by preserving spatial integrity.
• The method shows high potential for diverse real-world applications requiring robust data representation.
• The seamless integration of projection, representation, and manifold learning enhances analytical capabilities.