A Scoring Scheme for Online Feature Selection: Simulating Model Performance Without Retraining.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Increasing model complexity with irrelevant features can lead to overfitting and reduced interpretability.
• Online learning and real-world scenarios necessitate periodic model retraining or performance testing with newly discovered features.
• Current feature selection methods, supervised or unsupervised, present computational or subjective limitations.

Purpose of the Study:

• To introduce an accurate feature importance score for optimal feature selection.
• To address the challenges of online feature selection in dynamic environments.
• To enable performance evaluation of new features without model retraining.

Main Methods:

• Development of a novel scoring mechanism to determine feature importance.
• Application of the score in online feature selection scenarios.
• Evaluation of the score's ability to interpret performance improvements.

Main Results:

• The proposed score accurately determines feature importance.
• The score offers low time complexity suitable for online scenarios.
• Performance improvement can be interpreted without invoking model retraining.

Conclusions:

• The introduced score provides an efficient and effective solution for online feature selection.
• This method enhances model interpretability and reduces computational overhead.
• The score facilitates dynamic model adaptation in response to new data features.