Directional Clustering Through Matrix Factorization.

Area of Science:

• Machine Learning
• Data Mining
• Computational Neuroscience

Background:

• Clustering algorithms often rely on Euclidean distance, which may not capture directional relationships effectively.
• Existing methods like nonnegative matrix factorization (NMF), K-EVD, and k-means have limitations in handling angle-based clustering.
• Applications in document classification and brain imaging require grouping data based on directional similarity.

Purpose of the Study:

• To develop a novel clustering algorithm that effectively groups feature vectors based on their directional similarity (angle).
• To compare the proposed algorithm against established clustering techniques in terms of performance and computational efficiency.
• To demonstrate the algorithm's utility in diverse applications, including text analysis and neuroimaging.

Main Methods:

• The proposed method utilizes concepts from constrained low-rank matrix factorization and sparse approximation.
• It employs iterative matrix decomposition to optimize a data fidelity term without enforcing strict positivity constraints or explicit eigenvector computation.
• Each optimization step is followed by a hard cluster assignment, similar to k-means and K-EVD.

Main Results:

• The novel algorithm demonstrates superior clustering performance compared to common competitors.
• It achieves significant improvements in computational speed for various datasets.
• Empirical evaluations on toy problems, text datasets, and functional magnetic resonance imaging (fMRI) data validate its effectiveness.

Conclusions:

• The developed angle-based clustering algorithm offers an efficient and effective alternative to traditional methods.
• Its ability to handle directional data makes it suitable for complex applications like document classification and brain region partitioning.
• The approach provides a promising direction for future research in unsupervised learning and data analysis.