Effect of Outlier Detection Techniques in Machine Learning Assisted Mass Real Estate Valuation

In mass real estate valuation, detecting outliers is crucial for improving model performance. This study examined how different outlier detection techniques impact model performance when using machine learning algorithms. A dataset with 78 variables and 37,269 market samples, enriched with geographic dataset, was created for Tuzla and Pendik districts in Istanbul. Outliers were detected using Boxplot, MAD, and Z-Score techniques. The Boxplot detected the most outliers (3,810), while Z-Score (|Z| > 3) detected the fewest (943). After removing outliers, the dataset was split into training (70%) and test (30%) sets. Then, Random Forest (RF) models were trained, and performance was evaluated using MAE, MAPE, RMSE, and R² metrics. Methods have emerged from different perspectives. Z-Score (|Z| > 3) produced the best results in terms of absolute error and overall performance, achieving the lowest MAE (0.046849) and RMSE (0.065869). The Boxplot method showed consistent performance, with notably low MAPE (12.933) and strong explanatory power (R² = 0.698), making it a good general-purpose technique. MAD provides the highest R² (0.700), making it the most effective in explaining variance. These findings highlight that determining appropriate outlier detection methods can lead to more accurate and robust valuation models by optimizing datasets.