Impact of Filter-Based Feature Selection on Prediction Model Performance for Early Detection of Stunting

This study investigates the impact of feature selection results using the filter method on the performance of predictive models for the nutritional status of children aged 0-23 months. This study aimed to understand how correlation-based feature selection methods affect the efficiency and accuracy of predictive models. The dataset comes from the Basic Health Research survey, which consists of 84 variables and around 130,000 data, including anthropometric, family, socioeconomic, and environmental information. In this study, various filter-based feature selection methods were applied to evaluate and identify features that have significant relationships with children’s nutritional status, both linearly and non-linearly. The mutual information method was used to identify complex relationships between features and target variables, while the Pearson Correlation Coefficient and information gain were used to assess the relevance of numerical and categorical features. Of the 84 variables in the dataset, 37 features were selected after the selection process. The model performance evaluation was carried out using the Random Forest (RF) and Gradient Boosting (GB) algorithms, by comparing the results between the dataset that had been selected and without feature selection. The RF model without feature selection produces accuracy: 0,84600, precision: 0,83958, recall: 0,84600, F1-score: 0,78962, and a computation time of 9 seconds. The GB model without feature selection produces accuracy: 0,86330, precision: 0,84722, recall: 0,86330, F1-score: 0,84925, and a computation time of 2 minutes 50 seconds. After feature selection and cross-validation, the RF model shows accuracy: 0,85701, precision: 0,83599, recall: 0,85701, and F1 score: 0,83484. The GB model shows accuracy at 0,86124, precision at 0,84226, recall at 0,86130, and F1 score at 0,84120.