Feature Selection and Feature Stability Measurement Method for High-Dimensional Small Sample Data Based on Big Data...

Area of Science:

• Artificial Intelligence
• Data Mining
• Bioinformatics

Background:

• High-dimensional small sample data presents challenges like "dimensional disasters" in data mining.
• Existing feature selection methods often prioritize classification performance over stability, leading to unreliable feature identification.

Purpose of the Study:

• To develop a novel ensemble feature selection method that improves both classification performance and stability for high-dimensional small sample data.
• To analyze the causes of feature selection instability and introduce a metric for evaluating stability.

Main Methods:

• Proposed the Random Bits Forest Recursive Clustering Eliminate (RBF-RCE) ensemble feature selection method.
• Utilized parallel learning with multiple basic classifiers to optimize feature selection.
• Introduced the Intersection Measurement (IM) method to assess feature selection stability.

Main Results:

• The RBF-RCE method demonstrated improved classification performance compared to traditional methods.
• Experiments confirmed enhanced stability in feature selection results using the RBF-RCE approach.
• The Intersection Measurement (IM) effectively evaluated the stability of the proposed method.

Conclusions:

• The RBF-RCE method offers a robust solution for feature selection in high-dimensional small sample datasets.
• Improved stability in feature selection leads to more reliable and interpretable features.
• The proposed method contributes to more effective data mining in complex biological and computational fields.