Comparative Study of Feature Selection Techniques in Machine Learning Algorithms for Predicting Female School Attrition in Bangladesh

In real-world scenarios, such as school attrition, it is important to consider relevant features when applying machine learning algorithms. However, not all features are essential, and adding irrelevant ones can diminish the effectiveness of the algorithm. Feature selection (FS) is a vital process in machine learning to minimize data dimensionality and increase algorithm efficiency. Evolutionary algorithms (EA) have attracted plenty of attention as a technique of solving FS challenges. Surprisingly, no previous research has applied these methods specifically to predict factors related to female school attrition. Thus, the purpose of this study is to use multiple feature selection strategies. Genetic Algorithm (GA), Chi-squared, Mutual Information (MI), Exhaustive Feature Selection (EFS) and evaluate their effectiveness by comparing results with and without the application of feature selection techniques. Data for this study is collected from the Multiple Indicator Cluster Surveys, 2019. Among the methods tested, artificial neural networks (ANN) demonstrated the highest accuracy in predicting school attrition. Using GA, Chi-squared, MI, and EFS techniques, the ANN model’s accuracy was 83.76%, 84.11%, 84.23%, and 84.88%, respectively. Using the same feature selection methods, Support Vector Machine (SVM) produced results of 82.49%, 83.18%, 81.53%, and 82.65% whereas Logistic Regression (LR) produced accuracy numbers of 81.72%, 81.38%, 82.29%, and 82.31%. Without applying any feature selection techniques, ANN provided 82.38%, SVM provided 81.17% and LR provided 81.14% of accuracy. Therefore, the findings indicate that all models performed well in classifying students as in-school or out-of-school and we achieved higher accuracies after applying FS approaches than without FS. Thus, tailoring filter measures based on specific FS technique characteristics can significantly enhance efficiency and effectiveness.