House Pricing (Machine Learning)

Machine Learning project to estimate real estate prices from structured and semi-structured features.
The goal is to build a robust data preprocessing pipeline and compare different models to identify the solution with the best performance.

---

📌 Project contents

• Main notebook: includes
  • complete preprocessing (cleaning + feature engineering)
  • creation of *_clean.csv datasets
  • model training and comparison (Random Forest, Linear Regression, Lasso, Gradient Boosting)
  • evaluation on a holdout test set
• Dataset
  • train.csv, test.csv (raw)
  • train_clean.csv, test_clean.csv / or X_clean.csv, y_clean.csv (processed)
• Output
  • final prediction files (e.g., predicted_values_*.txt)

---

🧠 Approach

1) Data Preprocessing / Feature Engineering

The dataset contains numerical, categorical, and semi-structured (text-like) variables.
Preprocessing was designed to make the pipeline reproducible and reduce errors due to non-uniform formats.

Main transformations:

Parsing and normalization of semi-structured features

• car_parking: extract the number of parking spots, distinguishing (e.g.) garage vs shared parking and adding a missing flag.
• availability: handle values such as available or available from dd/mm/yyyy, transforming them into indicators and time components (year/month).
• condominium_fees: convert strings to numeric values, handling special cases (e.g., no fees).
• floor / total_floors_in_building: numeric parsing and special cases (e.g., ground floor, mezzanine).
• energy_efficiency_class: map to an ordinal scale and add a missing flag.
• year_of_construction: numeric conversion + derived feature building_age.

“Simple but effective” feature engineering

• is_top_floor, floor_ratio
• num_balconies extracted from other_features (e.g., "2 balconies")

Handling multi-label features (other_features)

• split on the | separator + fixes for known concatenation issues
• conversion to multi-hot encoding via MultiLabelBinarizer

Categorical encoding

• One-Hot Encoding for conditions and zone
• handle_unknown="ignore" to manage categories that appear only in the test set

Missing value imputation

• imputation using statistics computed on the training set (e.g., medians) to avoid leakage.

Result: creation of final clean, numeric datasets:

• train_clean.csv / test_clean.csv (or X_clean.csv / y_clean.csv)

---

2) Model training and comparison

After preprocessing, the data is split into:

• Train
• Holdout test (for a fair comparison across models)

Trained and compared models:

• Linear Regression (OLS) – linear baseline
• Lasso (LassoCV) – regularization + implicit feature selection
• Random Forest Regressor – non-linear model, robust to heterogeneous features
• Gradient Boosting Regressor – boosting to improve accuracy

For linear models, StandardScaler is used (within a pipeline) to ensure consistent scaling.

Metrics used on the test set:

• RMSE (root mean squared error)
• R²

Finally, the model with the lowest RMSE is selected as the “best” candidate.

---

✅ Results

The notebook generates a table with RMSE and R² on the holdout test set for each model, along with comparison plots.
The best model is automatically identified and (optionally) used to generate final predictions on the test dataset.

---

▶️ How to run

Requirements

Python 3.9+ recommended.

Install dependencies:

pip install -r requirements.txt