Hybrid Subset Feature Selection And Importance Framework

Feature selection algorithms are used in high-dimensional data to remove noise, reduce model overfitting, training and inference time, and get the importance of features. Features subset selection is choosing the subset with the best performance. This research provides a Hybrid subset feature selection and importance (HSFSI) framework that provides a pipeline with customization for choosing feature selection algorithms. The authors propose a hybrid algorithm in the HSFSI framework to select the best possible subset using an efficient exhaustive search. The framework is tested using the Bombay stock exchange IT index’s companies’ data collected quarterly for 16 years consisting of 71 financial ratios. The experimental results demonstrate that models created using 12 features chosen by the proposed algorithm outperform models with all features with up to 6% accuracy. The importance-based ranks of all features are generated using the framework calculated using 13 implemented feature selection techniques. All selected feature subsets are cross-validated using prediction models such as support vector machine, logistic regression, KNeighbors classffier, random forest, and deep neural network. The HSFSI framework is available as an open-source Python software package named ‘‘feature-selectionpy’’ available at GitHub and Python package index.