Are screening methods useful in feature selection? An empirical study.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Filter methods are common preprocessing steps to reduce variables for learning algorithms.
• An objective evaluation is needed to compare filter methods and assess their utility against learning algorithms alone.

Purpose of the Study:

• To objectively evaluate popular filter methods for variable selection in machine learning.
• To compare the performance of filter methods against each other and when used with various learning algorithms.

Main Methods:

• Ten real-world datasets were used for evaluation.
• Popular screening methods were paired with three regression and five classification learners.
• Accuracy metrics, including R-square and Area Under the ROC Curve (AUC), were employed.

Main Results:

• Filter methods demonstrated utility in improving prediction accuracy on specific datasets.
• Improvements were observed on two regression and two classification datasets out of ten.
• Comparative analysis through curve plots and tables highlighted method performance variations.

Conclusions:

• Screening methods can be beneficial for enhancing machine learning model performance in certain scenarios.
• The effectiveness of filter methods is dataset and learner dependent.
• Objective evaluation is crucial for understanding the contribution of preprocessing techniques.