A parameter free relative density based biclustering method for identifying non-linear feature relations.

Area of Science:

• Data Mining and Machine Learning
• Bioinformatics and Computational Biology

Background:

• Traditional biclustering algorithms often rely on restrictive assumptions like linearity or monotonicity.
• Existing density-based methods may miss biclusters due to global density criteria.

Purpose of the Study:

• To introduce PF-RelDenBi, a novel biclustering algorithm that identifies biclusters based on local feature density variations.
• To overcome limitations of existing methods by handling non-linear and non-monotonous feature relationships.
• To develop a parameter-free algorithm applicable across diverse datasets.

Main Methods:

• PF-RelDenBi utilizes local variations in marginal and joint densities for feature pairs to identify observation subsets.
• A non-linear feature relation index is employed to find feature sets connected by common observations, forming biclusters.
• The algorithm operates without requiring user-defined parameters.

Main Results:

• PF-RelDenBi demonstrated superior performance on most simulated datasets compared to eleven state-of-the-art algorithms.
• Biclusters detected on benchmark datasets improved classification performance when used as additional features.
• On three benchmark datasets, PF-RelDenBi achieved higher accuracy, NMI, and ARI than eleven comparative methods.
• Application to a COVID-19 dataset identified demographic features influencing disease spread.

Conclusions:

• PF-RelDenBi effectively identifies biclusters with non-linear and non-monotonous feature relations.
• The parameter-free nature and robust performance make it suitable for various data mining applications.
• The method shows promise in feature engineering for improved classification and in identifying factors influencing disease transmission.