Unsupervised hierarchical clustering using the learning dynamics of restricted Boltzmann machines

Area of Science:

• Computational biology
• Machine learning
• Data science

Background:

• Real-world datasets often exhibit complex hierarchical structures.
• Identifying these hidden data structures is crucial for various applications.

Purpose of the Study:

• To present a general, interpretable method for building relational data trees.
• To exploit the learning dynamics of restricted Boltzmann machines for structure discovery.

Main Methods:

• Utilized a mean-field approach derived from Plefka expansion within disordered systems.
• Applied the method to artificially generated and real-world datasets (digit images, genomic mutations, protein families).

Main Results:

• The method successfully and automatically identified the hierarchical structure in all tested datasets.
• Demonstrated effectiveness in uncovering complex relationships within biological data.

Conclusions:

• The developed method provides an effective tool for discovering hierarchical data structures.
• Potential applications include the study of homologous protein sequences for functional and evolutionary insights.