Estimation of Discriminative Feature Subset Using Community Modularity.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Feature selection (FS) is crucial for optimizing machine learning and data mining tasks.
• Existing methods may not adequately capture inter-feature relationships, leading to suboptimal subset selection.

Purpose of the Study:

• To propose a novel group-based feature selection method.
• To enhance the evaluation of feature subsets by focusing on inter-dependency rather than individual feature relevance.

Main Methods:

• Constructing sample graphs (SG) for different k-feature subsets.
• Applying community modularity to the sample graphs to evaluate feature groups.
• Introducing the k-features sample graph feature selection (FS) method.
• Validating the method using k-means clustering theory.

Main Results:

• The proposed method effectively identifies highly informative features as a group.
• It accurately measures relevant in-dependency among selected features, distinguishing from redundancy.
• The approach determines discriminative cues of feature subsets with maximal in-dependency.

Conclusions:

• The k-features sample graph feature selection method offers an effective approach to feature subset evaluation.
• This community modularity-based technique outperforms state-of-the-art methods in experiments.
• The method enhances preprocessing for machine learning by selecting interdependent features.