Budget constrained non-monotonic feature selection

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Feature selection is crucial in machine learning and data mining.
• Traditional methods exhibit a monotonic property, limiting effectiveness for non-monotonic feature dependencies.
• Budget constraints on feature subset size present a significant challenge.

Purpose of the Study:

• To develop an algorithm for non-monotonic feature selection.
• To address the limitations of traditional monotonic feature selection methods.
• To approximate the combinatorial optimization problem using Multiple Kernel Learning (MKL).

Main Methods:

• Developed a novel algorithm for non-monotonic feature selection.
• Approximated the combinatorial optimization problem via Multiple Kernel Learning (MKL).
• Provided performance guarantees for the derived solution compared to the global optimum.

Main Results:

• The proposed framework demonstrates promising performance on synthetic and real-world datasets.
• Empirical evaluations were conducted for both classification and regression tasks.
• The algorithm effectively handles non-monotonic feature selection under budget constraints.

Conclusions:

• The developed MKL-based approach offers an effective solution for non-monotonic feature selection.
• The framework outperforms baseline feature selection methods.
• This work advances feature selection techniques by addressing non-monotonicity and budget constraints.