Comparative Analysis of Feature Selection Methods in Classification of Thyroid and Heart Disease Dataset

High-dimensional data cause serious problems in the scalability and performance of machine learning algorithms. Feature selection is a preprocessing method that selects important features from high-dimensional data sets to improve the efficiency and accuracy of pattern recognition. This study uses two public datasets, namely Differentiated Thyroid Cancer Recurrence (dataset1) and Heart Disease Classification (dataset2). The stages of this research are preprocessing, feature selection, classification, and performance evaluation. The experiments used three feature selection methods, namely ReliefF, Lasso, and Random Forest. As for the classification using SGDClassifier. The results show that Random Forest consistently shows good performance, achieving high ROC area scores (0.99 for thyroid disease and 0.82 for Heart Disease) and effective classification error reduction. ReliefF also showed good performance, while Lasso had a less stable and lower performance than other methods. Unfiltered features generally show strong initial performance but are less stable than feature-selected models. Overall, in this study Random Forest was used as a feature selection method with stable performance due to its balanced performance in both datasets.