Feature Importance Analysis and Model Performance Evaluation for Real Estate Price Prediction

Real estate price prediction plays a significant role in decision-making for property investments. To enhance prediction accuracy and efficiency, in this research paper, we present a comprehensive analysis of real estate pricing data, with a focus on predictive modeling and feature engineering. We initially train several machine learning models on a dataset containing multiple property attributes, including Linear Regression, Lasso, Elastic Net, Support Vector Regression (SVR), Gradient Boosting, Random Forest, XGBoost, and Ridge. Observations highlight the superiority of tree-based models over linear-based ones, with Random Forest demonstrating the lowest Mean Absolute Error (MAE) of 37.744. Subsequently, feature selection and hyperparameter tuning revealed substantial MAE improvements in Lasso and Elastic Net with specific adjustments. It also emphasized the minimal impact of feature selection on SVR’s performance, even with only the top 10 features, and the notable reduction in MAE in the case of Lasso with the top 20 features. Furthermore, a new feature is introduced “distance from centre,” based on latitude and longitude coordinates, demonstrating its impact on MAE reduction. These findings underscore the importance of model-specific adjustments and feature engineering in enhancing predictive accuracy within the real estate domain.