Feature Interaction for Streaming Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Bioinformatics

Background:

• Traditional feature selection assumes complete data, unsuitable for data/feature streams.
• Existing streaming methods overlook feature interactions, treating features individually.
• Feature interactions are crucial as combined features can be more predictive than individual ones.

Purpose of the Study:

• To address the challenge of feature interaction in streaming feature selection.
• To propose a novel streaming feature selection method that considers feature interactions.
• To develop a metric for quantifying feature interaction in feature streams.

Main Methods:

• Introduced Streaming Feature Selection considering Feature Interaction (SFS-FI).
• Defined feature interaction formally.
• Designed a new metric, 'interaction gain', to measure feature interaction degree.
• Analyzed the relationship between feature relevance and feature interaction.

Main Results:

• The proposed SFS-FI method effectively selects interacting features.
• The 'interaction gain' metric accurately quantifies feature interaction.
• Experiments on 14 real-world microarray datasets demonstrate the method's efficiency.

Conclusions:

• Feature interaction is a critical aspect of streaming feature selection.
• The SFS-FI method offers an effective solution for handling feature interactions in streams.
• The developed 'interaction gain' metric advances the field of streaming feature selection.