Concept Factorization With Local Centroids.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Matrix factorization methods offer compact data representation for clustering.
• Existing methods often assume global centroids, failing with complex data structures.
• Current techniques may not ensure samples associate with the nearest centroid.

Purpose of the Study:

• To propose a novel data clustering approach, Concept Factorization with Local Centroids (CFLCs).
• To address limitations of existing methods in handling complex data structures and centroid association.
• To improve the interpretability and accuracy of data clustering.

Main Methods:

• Developed the CFLC model allowing samples to connect with multiple local centroids.
• Modeled pairwise relationships between samples and centroids for accurate label assignment.
• Formulated clustering as a bipartite graph partitioning task with an efficient optimization algorithm.

Main Results:

• The CFLC model successfully captures manifold structures in data.
• Experimental results demonstrate superior performance compared to state-of-the-art methods.
• Validated the effectiveness of CFLCs on diverse datasets.

Conclusions:

• CFLCs offers an effective solution for data clustering, especially for complex datasets.
• The model's ability to capture local structures and improve label assignment is a key advantage.
• CFLCs represents a significant advancement in matrix factorization-based clustering techniques.