DDVFA

DDVFA implements unsupervised clustering within the Adaptive Resonance Theory (ART) framework by combining global ART systems and local fuzzy ART modules to learn arbitrarily-shaped clusters incrementally and mitigate order dependence.

Key Features:

• Modular Architecture: A distributed architecture where each node is a local fuzzy ART module enabling multi-prototype cluster representations.
• Dual Vigilance Mechanism: Dual vigilance controls allow adaptive sensitivity to input patterns to reduce category proliferation and refine cluster granularity.
• Higher-Order Activation and Match Functions: Distributed higher-order activation and match functions support management of complex data structures and improved clustering accuracy.
• Order Dependence Mitigation (VAT pre-processing): Pre-processing with Visual Assessment of Cluster Tendency (VAT) for batch mode to reveal inherent cluster structures before learning.
• Order Dependence Mitigation (Merge ART module): A Merge ART post-processing module that refines clusters online as new data are introduced.
• Online and Offline Learning Capabilities: Supports incremental (online) learning with random sample presentation and batch (offline) learning when pre-processed with VAT.
• Performance Parity with Non-ART Methods: Experimental results show statistical equivalence to non-incremental algorithms such as DBSCAN, SL-HAC, and k-means while retaining ART incremental advantages.

Scientific Applications:

• Real-time genomics and proteomics analysis: Incremental learning suits continuous or streaming data in genomics and proteomics.
• Clustering of complex biological datasets: Handles arbitrarily-shaped clusters and reduces order dependence for complex bioinformatics datasets.

Methodology:

Combination of global ART systems and local fuzzy ART modules with distributed higher-order activation and match functions, a dual vigilance mechanism, VAT pre-processing for batch mode, and a Merge ART module for online cluster refinement.