A Novel Approach to Feature Selection Using Correlation Coefficients and Weighted Metrics

The consequences and efficiency of machine learning (ML) and data mining (DM) algorithms explicitly rely on the data's benchmark. Irrelevant or redundant data negatively hinders the model from achieving the preferred performance while simultaneously creating additional computational overhead; hence, models employ distinct data processing techniques. The feature selection (FS) technique is one of the best data preprocessing methods for selecting the most influential features while maintaining the model's stability, optimizing its performance through meticulous use, and enabling rapid training to provide accurate predictions. Accordingly, this article proposes an FS method based on the correlation coefficient and weighted metrics (FS-C2WM), which can moderately reduce the number of extrinsic features and enhance the model's predictive outcomes by providing an optimistic approach. Ten diverse datasets were utilized for experimental analysis, and the k-nearest neighbors (KNN) algorithm was employed to assess the proposed model's performance. A noticeable advancement is observable in the experimental results of FS-C2WM compared to the other prominent FS techniques and no selection (NoSel). FS-C2WM demonstrated a superior exhibition of mean values, outperforming the NoSel and other FS algorithms in accuracy, precision, recall, and f-score by at least 1.46%, 1.79%, 1.46%, and 1.16%, respectively. Findings reveal that the presented method can boost predictive performance, open new horizons by simplifying mathematical complexity, assemble top-tier classification models, and enhance the interpretability of ML models by removing features.