Grab'Em: A Novel Graph-Based Method for Combining Feature Subset Selectors.

Area of Science:

• Data Mining and Machine Learning
• Computational Statistics

Background:

• Feature selection is crucial for improving classification performance and understanding data mining problems.
• Combining results from multiple feature selection processes, such as ensembles or data division, is common.
• Existing methods often use simple voting based on selection frequency, neglecting feature interdependencies.

Purpose of the Study:

• To propose a new feature selection approach that accounts for feature relationships.
• To overcome the limitations of standard voting schemes in combining feature selection results.
• To enhance the effectiveness of feature selection in ensemble and data division methods.

Main Methods:

• Developed a novel approach using feature co-occurrence counts instead of simple selection frequencies.
• Constructed an undirected graph where vertices represent features and edges represent co-selection frequency.
• Utilized this graph to select feature subsets, explicitly avoiding redundancy.

Main Results:

• The proposed graph-based method outperforms the standard voting scheme.
• Redundancy issues inherent in simple voting are mitigated.
• Improved results were observed in both ensemble and data division scenarios for feature selection.

Conclusions:

• Considering feature relationships through co-occurrence analysis is superior to simple frequency-based voting.
• The graph-based approach offers a more robust method for combining feature selection results.
• This technique enhances the quality of selected feature subsets in complex data mining tasks.