README.md

metadata

title: Application1
emoji: 🐠
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 6.9.0
app_file: app.py
pinned: false
short_description: Appartement price prediction

Zurich Apartment Rent Predictor

This application predicts monthly rental prices (CHF) for apartments in the Canton of Zurich.

Live App: TODOOOOOOOOOHugging Face Spaces – Zurich Apartment Predictor

Model Iterations

Summary Table

Iteration	Objective	Key Changes	Models Used	CV Mean R²	CV Std	Δ Performance	Fit Diagnosis
1	Baseline	Basic cleaning, no log transform, 9 features, default hyperparameters	Linear Regression, Random Forest (n=100)	0.61 (RF) / 0.60 (LR)	0.10 / 0.06	Baseline	Overfitting (RF)
2	Improve generalization	Log-transform target, 32 features incl. income_density_score (new), lat/lon, Kreis dummies, tuned hyperparameters	Ridge (α=10), RF (n=300, depth=15), Gradient Boosting	0.70	0.04–0.07	+0.09	Good Fit

Iteration 1 Baseline

Establish a performance baseline.

Steps:

• Loaded apartments_data_enriched_with_new_features.csv
• Dropped rows with missing values in selected columns
• Used raw (untransformed) price as target
• Evaluated with 5-fold cross-validation (KFold, shuffle=True, random_state=42)

Models and Hyperparameters:

• LinearRegression
• RandomForestRegressor(n_estimators=100, random_state=42)

Features: rooms, area, pop_dens, frg_pct, tax_income, luxurious, furnished, temporary, zurich_city

Results:

Model	CV R²	CV Std	CV RMSE (CHF)
Linear Regression	0.60	0.06	~679
Random Forest	0.61	0.10	~666

Random Forest shows large gap between train R² (0.93) and test R² (0.61) → Linear model underfits. Both need improvement.

---

Iteration 2 Improved Model

Reduce overfitting, improve generalization with richer features and better hyperparameters.

Steps:

• Log-transform target variable to reduce right skew and improve linear model performance
• Engineered new feature: income_density_score = (tax_income × pop_dens) / 1,000,000
• Engineered additional features: log_area, rooms_area_ratio
• Added geographic features: lat, lon
• Added Zurich district dummies: Kreis 1–12
• Added keyword flags from listing text: (LOFT),(SEESICHT), etc.

Models and Hyperparameters:

• Ridge(alpha=10.0)
• RandomForestRegressor(n_estimators=300, max_depth=15, min_samples_leaf=3, random_state=42)
• GradientBoostingRegressor(n_estimators=300, max_depth=5, learning_rate=0.05, random_state=42)

Features: rooms, area, log_area, pop_dens, frg_pct, tax_income, income_density_score, rooms_area_ratio, lat, lon, luxurious, furnished, temporary, zurich_city, room_per_m2, Kreis 1–12, (ATTIKA), (LOFT), (SEESICHT), (LUXURIÖS), (POOL), (EXKLUSIV)

Results:

Model	CV R²	CV Std
Ridge (alpha=10)	0.70	0.04
Random Forest (tuned)	0.70	0.07
Gradient Boosting	0.70	0.07

All three models achieve ~0.70 R². RF train/test gap reduced to ~0.10 (was ~0.32). Good Fit.

Iteration 3 Room Scaling and Correct Defaults

Fix the prediction plateau and improve scaling across all room counts and price ranges.

Steps:

• Added log_rooms
• Added area_rooms_interact = area × rooms
• Corrected pop_dens defaults in the app: 4,729 for Zurich City, 1,328 for outside (both derived from training data medians)

Models and Hyperparameters:

• Ridge(alpha=10.0)
• RandomForestRegressor(n_estimators=300, max_depth=15, min_samples_leaf=3, random_state=42)
• GradientBoostingRegressor(n_estimators=800, max_depth=5, learning_rate=0.03, min_samples_leaf=3, subsample=0.8, random_state=42)

Features: rooms, area, log_area, log_rooms, pop_dens, log_pop_dens, frg_pct, tax_income, income_density_score, log_income_dens, rooms_area_ratio, area_rooms_interact, lat, lon, luxurious, furnished, temporary, zurich_city, room_per_m2, Kreis 1–12, (ATTIKA), (LOFT), (SEESICHT), (LUXURIÖS), (POOL), (EXKLUSIV)

Results:

Model	CV R²	CV Std
Ridge (alpha=10)	0.72	0.04
Random Forest (tuned)	0.72	0.05
Gradient Boosting (tuned)	0.72	0.05

Example predictions after fix (Zurich City, Kreis 3):

Rooms	Area	Predicted Rent
1.0	30 m²	CHF 1,157
3.0	80 m²	CHF 3,121
4.5	115 m²	CHF 3,741
5.5	145 m²	CHF 4,972
6.5	175 m²	CHF 5,879

---

New Feature: income_density_score

income_density_score = (tax_income × pop_dens) / 1,000,000

This feature combines municipal tax income with population density to create a neighbourhood affluence proxy. A dense area with high-income residents (for example Kreis 1 or Küsnacht) will score much higher than a sparse rural municipality.

Final Selected Model

Model: RandomForestRegressor(n_estimators=300, max_depth=15, min_samples_leaf=3)
Target: log(1 + price) → predictions are back-transformed with expm1()
CV R²: 0.70 | CV Std: 0.07
Evaluation: 5-fold cross-validation

Top Feature Importances:

1. log_area — log-transformed living area
2. area — living area in m²
3. rooms — number of rooms
4. income_density_score (new feature)
5. lat — latitude (geographic location)