Apartment Price Prediction – Regression Project

Project Overview

This project focuses on predicting apartment prices in a specific European region using advanced machine learning techniques. The dataset contains detailed apartment attributes, including size, number of rooms, floor information, building year, distances to nearby amenities, and neighborhood characteristics. The goal is to build a robust regression model that accurately estimates apartment prices (price_z) based on these predictors.

This regression problem reflects a real-world scenario, where accurate property valuation is critical for buyers, sellers, real estate agencies, and urban planners. The developed model can provide quick, data-driven price estimates, account for local market dynamics, and assist in investment or policy decisions.

Dataset

The dataset consists of:

• Training set: 156,454 observations with 34 features + target (price_z)
• Test set: 39,114 observations with 33 features (target not included)

Key features include: apartment size (dim_m2), number of rooms (n_rooms), building characteristics, distances to amenities, ownership type, condition, and additional synthetic indicators such as market_volatility, infrastructure_quality, popularity_index, and green_space_ratio.

Project Structure

The project is organized to reflect a professional workflow: ├── notebooks/ # Separate notebooks for preprocessing, EDA, model building, validation, and testing ├── data/ # Raw and processed datasets ├── src/ │ └── utils/ # Helper functions and pipeline utilities ├── visualizations/ # Plots and charts for EDA and model evaluation ├── models/ # Saved models using pickle ├── pipelines/ # Preprocessing and modeling pipelines ├── README.md └── .gitignore

Version control was managed using Git, and large model computations were performed in Google Colab. Local development and testing were done using VS Code and Jupyter Notebook.

Data Preprocessing

• Missing values handled using KNN imputation.
• Target variable (price_z) was log-scaled to normalize distribution.
• Features standardized using StandardScaler in pipelines.
• Categorical variables encoded for model compatibility.
• Dataset split: 50% train / 25% validation / 25% test to ensure proper evaluation.

Modeling Approach

The project followed a structured machine learning workflow:

1. Pipeline Creation – Preprocessing + model steps were combined into pipelines for reproducibility.
2. Model Exploration – Base algorithms included:
   • Linear Regression, Ridge, Lasso
   • Support Vector Regression (SVR)
   • Polynomial Regression
   • Random Forest
3. Hyperparameter Tuning – Applied GridSearchCV and RandomizedSearchCV for algorithm optimization:
   • Polynomial degree tuning
   • Regularization (Lasso/Ridge)
   • SVR kernel and C parameters
   • Tree depth, number of estimators, and subsample rates
4. Ensemble Methods – To boost predictive performance:
   • Boosting: XGBoost, LightGBM, CatBoost (gradient boosting to reduce bias)
   • Bagging: Random Forest, Bagged Decision Trees (variance reduction)
   • Stacking: Meta-model combining XGBoost, LightGBM, CatBoost, and Random Forest for superior accuracy
5. Model Evaluation – Metrics computed: RMSE, MAE, Median AE, MAPE, R².
6. Model Persistence – Final models, preprocessed data, and selected features saved using pickle for production deployment.

Validation and Test Results

During validation, multiple models were compared. The meta-stacked model combining boosting and bagging methods outperformed all individual models.

Test Set Performance (Stacked Meta-Model):

Metric	Value
RMSE	0.1091
MAE	0.0925
MedAE	0.0879
MAPE	0.6895
R²	0.9525

• The stacked model slightly outperformed XGBoost, LightGBM, and CatBoost individually, achieving 95% explained variance.
• This demonstrates that combining multiple models via stacking can capture complementary patterns in the data, improving prediction stability and accuracy.

Individual Boosting Model Example (XGBoost):

Metric	Value
RMSE	0.1094
MAE	0.0926
MedAE	0.0880
MAPE	0.6900
R²	0.9523

Tools and Technologies

• Python: Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn
• Modeling: Linear models, Polynomial Regression, Ridge/Lasso, SVR, Random Forest, Boosting (XGBoost, LightGBM, CatBoost), Stacking
• Hyperparameter Tuning: GridSearchCV, RandomizedSearchCV
• Persistence: Pickle
• Version Control: Git & GitHub
• Computation: Google Colab for heavy computations, local VS Code/Jupyter for development

Future Work

• Apply additional ensemble techniques (Gradient Boosting variations, Voting Regressor).
• Analyze feature importance from boosting and bagging models for interpretability.
• Include time-dependent economic indicators for seasonal or market trend capture.
• Deploy as production-ready service with real-time predictions and API integration.

Real-World Impact

This regression model can serve as a reliable price estimation tool for the real estate market:

• Quickly estimate apartment prices based on multiple predictors.
• Understand the influence of location, amenities, and apartment characteristics on price.
• Support data-driven investment, sales, or policy decisions.
• Stack-based meta-model ensures high reliability and robustness, suitable for production deployment in applications or dashboards.

---

This project demonstrates an end-to-end machine learning solution, from preprocessing and model building to ensemble optimization and production-ready deployment, following best practices for professional ML workflows.