Enhanced Gradient based Optimizer with Random Forest Stacking Ensemble for Predicting Screen Time Behavioral Patterns in Indian Children: A Hybrid Machine Learning Approach

The growing rate of exposure to excessive screen time in children in India has also increased the demand of predictive models that can be used reliably to determine the patterns of behavior risks at an early age. This paper presents a proposal of an Enhanced Gradient-Based Optimizer combined with a Random Forest stacking ensemble to classify behavioral risk levels of the screen time behavior by utilizing a set of behavioral and demographic features together. It is a research based on an 8,742 records of children between the ages of 4 to 16 years, including 24 variables that determine the screen use habits, physical activities, sleep habits, academic variables, and the socio demographic variables. The systematic cleaning, normalization and encoding of data were applied and feature selection based on mutual information was done to drop the less discriminative features. Synthetic Minority Over-sampling Technique was used to target the issue of class imbalance and allow equal learning opportunities in terms of risk category. The proposed stacking model incorporates the use of Random Forest, Extra Trees, Gradient Boosting, and XGBoost as base learners and a Random Forest classifier as the meta learner. RandomizedSearchCV was used to optimize hyper parameters on the basis of weighted F1-score as the objective criterion. Accuracy, precision, recall, F1-score, and multi-class ROCAUC were used as model performance metrics, and compared across performance on Decision Tree, Random Forest, Support Vector machine and single-model XGBoost. The experimental outcomes indicate that the suggested model obtained the following results: its accuracy is$\mathbf{7 3. 9}$percent, precision is$\mathbf{0. 7 5 4}$, recall is 0.739, an F 1 -score is 0.734, and the ROC-AUC is 0.801, which is an increase between 6 and 18 percentage points compared to the traditional methods. The analysis of stability based on box plots and violin plots revealed a lower variance and predictive behavior and a high correlation of the accuracy and$F 1$-score of 0.98, availed the balanced model performance. These results suggest that the suggested framework offers a strong and analytically justified framework of assessment of behavioral risk of screen time usage, and it could be potentially applicable to child health monitoring and intervention planning based on data.