prclust

prclust implements penalized regression–based clustering via the DC-ADMM algorithm, combining difference-of-convex (DC) programming and the alternating direction method of multipliers (ADMM) to provide scalable unsupervised clustering with finite-sample mis-clustering error bounds under L0-constrained regularization.

Key Features:

• DC-ADMM algorithm: Combines difference-of-convex programming with the alternating direction method of multipliers to perform penalized regression–based clustering.
• Closed-form updates: Uses closed-form updating formulas within the DC-ADMM framework to enhance computational efficiency.
• Scalability: Designed for large datasets to enable scalable computation in high-volume bioinformatics applications.
• Clustering consistency theory: Provides a theoretical framework including a finite-sample mis-clustering error bound with L0-constrained regularization.
• Versatile penalty support: Supports various loss functions and grouping penalty functions to tailor clustering to different data characteristics.
• Comparison to prior work: Demonstrates computational advantages relative to the quadratic penalty–based algorithm of Pan et al. (2013).

Scientific Applications:

• Genomics: Applicable to clustering tasks in genomic studies involving large-scale molecular measurements.
• Proteomics: Suited for grouping and exploratory analysis of proteomics datasets.
• Unsupervised exploratory analysis: Supports unsupervised exploratory clustering of large bioinformatics datasets with complex structure.

Methodology:

Implements penalized regression–based clustering via the DC-ADMM algorithm that integrates difference-of-convex programming and ADMM, employs closed-form update formulas, supports various loss and grouping penalties, and incorporates L0-constrained regularization with finite-sample mis-clustering error analysis.