Resampling-based similarity measures for high-dimensional data.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Assessing sample similarity is crucial for classification tasks.
• High-dimensional and megavariate datasets pose challenges for conventional similarity measures.
• Many features in large datasets may not be informative for sample differentiation.

Purpose of the Study:

• To propose a novel distance measure for assessing sample similarity in high-dimensional datasets.
• To address the limitations of traditional similarity metrics in complex biological data.
• To introduce an enriched feature selection strategy for improved similarity assessment.

Main Methods:

• A refiltering process involving random feature subset selection.
• Performing cluster analysis on each random feature subset.
• Calculating similarity based on the relative frequency of samples clustering together.
• Developing an enriched version with a higher probability for informative features.

Main Results:

• The proposed refiltering-based similarity measure demonstrates superior performance.
• The enriched form of the similarity measure is particularly effective.
• Outperforms conventional measures like Euclidean distance, correlation, and Hamming distance.
• Validated using real-world genomics datasets.

Conclusions:

• The refiltering-based similarity measure is a robust alternative for high-dimensional data.
• Enriching feature selection enhances the accuracy of sample similarity assessment.
• This method offers significant advantages for classification in genomics and related fields.