Automated Feature Reduction in Machine Learning
David A. Shilane · 2022 IEEE 12th Annual Computing and Communication Workshop and Conference (CCWC) · 2022
Supervised machine learning models are developed by selecting features that can be related to a dependent variable. This selection may involve a mix of automated procedures and manual exploration. Large data sets often include many features that are not suitable for the intended model. This paper proposes a framework for automated feature reduction (AFR) to identify features that can be removed from consideration for the model. It applies exclusion criteria based on the form of the data to determine whether each feature would pose practical issues for the design of the model. AFR is proposed as a preprocessing step prior to the use of automated or manual feature selection techniques. By reducing the set of possible features, this exploration can search a less complex space of potentially relevant features. AFR can be widely applied to many machine learning models and demonstrates clear benefits in improving the efficiency of feature selection algorithms. This paper provides details on the AFR algorithm along with examples of feature reduction in model development.