Local Sparse Bump Hunting.

Area of Science:

• Computational statistics
• Data mining
• Bioinformatics

Background:

• Identifying structures like clusters or classes in data is crucial for scientific discovery.
• Existing bump hunting methods struggle with high-dimensional and correlated data.
• Supervised and multivariate approaches are needed for complex datasets.

Purpose of the Study:

• To introduce a novel supervised, multivariate bump hunting strategy.
• To address challenges of correlation, interpretability, and high-dimensionality (p ≫ n).
• To enable class discovery with minimal assumptions.

Main Methods:

• A divide and conquer strategy combining tree-based methods, dimension reduction, and Patient Rule Induction Method (PRIM).
• Development of a method to estimate PRIM meta-parameters.
• Utilized cross-validation and ROC analysis for accuracy evaluation.

Main Results:

• The proposed method outperforms naive PRIM and competitive non-parametric methods in class discovery.
• Demonstrated effectiveness in identifying tumor subtypes in a colon cancer microarray dataset.
• Successfully handles noisy, high-throughput data.

Conclusions:

• The novel supervised bump hunting strategy is effective for exploring modes and classes in high-dimensional data.
• The method offers practical applications, particularly in analyzing complex biological datasets.
• Provides a robust tool for identifying underlying structures and facilitating scientific insights.