Jabodetabek House Price Prediction

House prices in Jabodetabek are influenced by many factors, including land size, building size, bedrooms, bathrooms, location, and property characteristics. The challenge was to build a regression model that could learn these relationships from real-world property listing data while dealing with missing values, categorical variables, and outliers. The project used several regression algorithms to understand which approach could provide the most reliable predictions.

I worked on the data preprocessing and Random Forest pipeline. The workflow included: • Cleaning missing and invalid data • Separating numerical and categorical features • Encoding property condition, certificate, furnishing, and city • Analyzing feature correlations • Handling outliers • Splitting the dataset into training and testing sets • Training multiple regression models • Evaluating models using MAE, MSE, RMSE, and R² • Tuning Random Forest hyperparameters using GridSearchCV After preprocessing and outlier handling, the dataset contained 3,423 clean records ready for model training.

For the Random Forest model, I compared several configurations: 1. Default Random Forest 2. Random Forest with hyperparameter tuning 3. Random Forest with feature scaling 4. Random Forest with selected features Hyperparameter tuning was performed using GridSearchCV with 5-fold cross-validation. The selected-feature approach focused on features such as: • Bedrooms • Bathrooms • Land size • Building size • Carports • Maid bedrooms • Maid bathrooms • Garages • City encoding Feature selection produced the strongest Random Forest result, achieving an R² Score of 0.883476.

The experiments showed that preprocessing strategies can have different effects depending on the algorithm. Random Forest with selected features achieved the best result with: • MAE: Rp641.2M • MSE: 2.4132 × 10¹⁸ • RMSE: Rp1.553B • R² Score: 0.883476 This was higher than the other evaluated models, including Gradient Boosting, KNN, and Linear Regression. The experiment also showed that normalization was particularly useful for distance-based algorithms such as KNN, while feature selection provided a noticeable improvement for tree-based models.