Evaluating Machine Learning Models for House Price Prediction with Different Sampling Techniques

Abstract

This study investigates the interplay between advanced sampling techniques and machine learning models to predict residential property sale prices using a diverse dataset encompassing structural, locational, and economic attributes. Emphasizing Stratified Extreme Ranked Set Sampling (SERSS), the research systematically evaluates the impact of five sampling methods—SERSS, Cluster, Bootstrap, Systematic, and Random Sampling—on various machine learning algorithms, including CatBoost, Random Forest, ElasticNet, and FIkNN. The findings reveal that SERSS significantly enhances the generalizability and robustness of predictive models by capturing both central and extreme data tendencies, outperforming traditional methods in preserving dataset variability. Ensemble methods like CatBoost, Random Forest and similarity algorithm like FIkNN consistently demonstrated superior predictive accuracy, achieving the Mean Absolute Error (MAE) between $85 and $650, and high R² values across structured sampling techniques. Conversely, unstructured methods such as Random Sampling introduced biases, leading to substantial deviations in predictions. These results underscore the critical importance of aligning sampling methodologies with model-specific characteristics to optimize performance. This study provides actionable insights for researchers and practitioners in predictive modeling, offering a framework for integrating sampling strategies with advanced machine learning models to tackle heterogeneous datasets effectively.