A Cheap Feature Selection Approach for the K-Means Algorithm.

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• High-dimensional data analysis poses challenges for K-means clustering.
• Existing dimensionality reduction techniques for K-means are often computationally expensive and difficult to parallelize.

Purpose of the Study:

• To develop a fully parallelizable feature selection technique for the K-means algorithm.
• To introduce a novel feature relevance measure tied to K-means error.

Main Methods:

• Propose a parallelizable feature selection technique based on a new relevance measure.
• Partition features, cluster each subset, and select features with the highest relevance.
• Analyze theoretical quality and empirical performance.

Main Results:

• The proposed method achieves lower K-means error than traditional feature selection techniques (Laplacian scores, max variance, etc.).
• It requires similar or lower computational times compared to existing feature selection methods.
• Demonstrates noticeable improvements in both error and computational time over feature extraction methods like random projections.

Conclusions:

• The novel feature selection technique offers an efficient and effective solution for high-dimensional K-means clustering.
• It provides a competitive alternative to existing feature selection and extraction methods, especially in terms of accuracy and speed.