Infinite Feature Selection: A Graph-based Feature Filtering Approach

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Feature selection is crucial for model performance and interpretability.
• Existing methods often struggle with complex feature interdependencies and scalability.
• Identifying optimal feature subsets remains a significant challenge in data analysis.

Purpose of the Study:

• To propose a novel filtering feature selection framework named Infinite Feature Selection (Inf-FS).
• To develop a method that handles feature relevance and redundancy within a unified graph-based approach.
• To enable efficient and effective feature subset selection, even with an infinite number of feature subsets.

Main Methods:

• Representing feature subsets as paths in a graph where nodes are features and edges represent their relations.
• Evaluating path values using matrix power series properties or Markov chain fundamentals.
• Developing an unsupervised strategy for determining the optimal feature subset size from ranked features.

Main Results:

• Inf-FS demonstrates superior performance compared to 18 widely-known feature selection approaches across 11 diverse benchmarks.
• The framework effectively handles heterogeneous features and various feature subset cardinality scenarios.
• Infinite evaluation allows for constrained computational complexity and elegant feature ranking.

Conclusions:

• Infinite Feature Selection (Inf-FS) offers a robust and efficient solution for feature selection.
• The graph-based approach effectively models feature interdependencies, improving selection accuracy.
• Inf-FS provides a scalable and high-performing alternative for diverse machine learning applications.