Published Paper Details:

Quality and relevance of features used in training are key factors in machine learning models' efficacy and dependability. Overfitting, higher computational complexity, and worse predictive performance can result from the redundant or unnecessary features that are frequently present in high-dimensional datasets. This work suggests a thorough architecture for feature selection that blends embedding, filtering, and wrapping methods to find the most important predictors while reducing noise. For the best subset selection, wrapper-based recursive feature elimination (RFE) is used after filter techniques like correlation analysis and chi-square tests are used to eliminate redundant features. To guarantee robust selection during model training, embedded strategies are employed, such as Lasso and tree-based feature significance algorithms. The suggested approach increases accuracy across a variety of machine learning techniques, decreases training time, and improves model interpretability. Comparing experimental outcomes to baseline models with all characteristics, and the F1-score show notable increases. This study highlights how crucial systematic feature selection is as a first step in creating scalable and effective predictive modelling pipelines.