Feature Selection Approach Based on Stacked Density Granulation With Principle of Justifiable Granularity

Area of Science:

• Machine Intelligence
• Data Mining
• Artificial Intelligence

Background:

• Information granularity aids machine intelligence in problem-solving and decision-making.
• Traditional methods like fuzzy c-mean (FCM) and K-means struggle with nonconvex data structures.
• Existing methods can lose data wholeness and accuracy when handling complex datasets.

Purpose of the Study:

• To propose a novel density-based clustering method for constructing robust information granularities.
• To introduce a feature selection method leveraging these granularities for improved performance.
• To address the limitations of existing information granularity techniques.

Main Methods:

• Developed stacked density granulation (SDG), a density-based clustering approach.
• Introduced density granular feature selection (DGFS) using SDG for feature importance measurement.
• Aggregated and discretized information granularities to create a low-dimensional feature space.

Main Results:

• SDG effectively describes both convex and nonconvex data, overcoming limitations of traditional methods.
• DGFS successfully identified and retained the most relevant features.
• Experiments on 12 datasets showed DGFS consistently outperformed other feature selection methods across four classifiers.

Conclusions:

• The proposed SDG and DGFS methods offer a superior approach to information granularity and feature selection.
• DGFS demonstrates significant effectiveness and robustness in enhancing machine intelligence capabilities.
• The findings suggest a promising direction for adaptive and flexible machine decision-making in complex environments.