IMPACT OF STACKING ENSEMBLE DEPTH ON GENERALIZATION ABILITY OF ACADEMIC PERFORMANCE PREDICTION MODELS

Abstract

Purpose. The research is aimed at a comprehensive analysis of the impact of stacking ensemble depth on the generalization ability of academic performance prediction models and determining the optimal stacking depth to achieve maximum performance and reliability of predictions. The goal of the work is to develop a methodology for assessing the relationship between stacking ensemble depth and model generalization metrics, as well as determining recommendations for selecting optimal ensemble architecture for academic performance prediction tasks.

Methodology. The research methodology is based on experimental analysis of the performance of stacking ensembles of different depths (from 1 to 5 levels) for predicting student academic performance. Base models include logistic regression, Random Forest, Gradient Boosting, Support Vector Machine, and neural networks. Generalization ability assessment is performed using accuracy, F1-score, AUC-ROC, and coefficient of determination metrics on independent test samples. Stratified cross-validation is applied to assess result stability and analyze the impact of stacking depth on model variance and bias.

Findings. Experimental results demonstrate a non-trivial relationship between stacking ensemble depth and model generalization ability. For single-level stacking (depth 1), generalization ability is 0.82 by F1-score metric, for two-level stacking (depth 2) – 0.87, for three-level (depth 3) – 0.89, for four-level (depth 4) – 0.88, for five-level (depth 5) – 0.86. The optimal stacking depth is identified at level 3, where maximum generalization ability is achieved without significant increase in model complexity. At depths greater than 3 levels, a decrease in generalization ability is observed due to error accumulation and meta-model overfitting. It is established that stacking depth affects the balance between model bias and variance, with optimal depth ensuring minimum generalization error.

Originality. A comprehensive assessing classification model stability depending on training sample size is developed, including theoretical analysis of the relationship between sample size and variance component of generalization error, empirical methods for determining saturation points, and comparative analysis of the effectiveness of different stability improvement methods. The impact of class imbalance and feature space dimensionality on the relationship between sample size and model stability is systematically investigated for the first time. A classification of models by dependence on training sample size is developed, taking into account algorithm type, model complexity, and data nature..

Practical value. The obtained results allow justifying the choice of optimal stacking ensemble depth for academic performance prediction tasks, ensuring high prediction accuracy with minimal model complexity. The developed recommendations can be applied in educational process management systems, early detection systems for at-risk students, and adaptive educational platforms. Determining optimal stacking depth allows optimizing the use of computational resources and ensuring high prediction reliability in practical applications.