README.md

metadata

title: Exercise1
emoji: 🏃
colorFrom: gray
colorTo: gray
sdk: gradio
app_file: app.py
pinned: false

Model Iterations Documentation

Task: Apartment Price Prediction (Regression)

Application Link

Public URL (Hugging Face Space):

https://huggingface.co/spaces/nbacchi/exercise1

---

Summary of Iterative Process

Iteration	Objective	Key Changes	Models Used	CV Mean R²	CV Std Dev	Change in Performance	Fit Diagnosis
1	Build baseline model	- Drop missing values - Remove duplicates - Price filter (750–8000 CHF) - Valid rooms/area filter - 5-fold CV	Linear Regression Random Forest (n_estimators=300)	0.5446 (LR) 0.5178 (RF)	0.1071 (LR) 0.1195 (RF)	Baseline	☑ Overfitting ☐ Underfitting ☐ Good Fit
2	Improve generalization	- Feature engineering - municipality_area_proxy = pop/pop_dens - emp_per_resident = emp/pop - foreigner_count_est = pop×(frg_pct/100) - Hyperparameter tuning - 5-fold CV	Ridge (alpha=1.0) Tuned Random Forest (n_estimators=500, max_depth=12, min_samples_split=5, min_samples_leaf=2)	0.5297 (Ridge) 0.5509 (RF)	0.0947 (Ridge) 0.1060 (RF)	+0.0331 (RF)	☐ Overfitting ☐ Underfitting ☑ Good Fit

---

Detailed Metrics Comparison

Iteration 1 – Baseline

Model	CV Mean R²	CV Std R²	CV Mean RMSE	CV Mean MAE
Linear Regression	0.5446	0.1071	673.00	468.07
Random Forest	0.5178	0.1195	698.51	500.13

Iteration 2 – Feature Engineering

Model	CV Mean R²	CV Std R²	CV Mean RMSE	CV Mean MAE
Ridge	0.5297	0.0947	682.01	481.08
Tuned Random Forest	0.5509	0.1060	674.54	473.98

---

Created Features

Iteration 2 Feature Engineering:

• municipality_area_proxy = population / population density
• emp_per_resident = employees / population
• foreigner_count_est = population × (foreigner_pct / 100)

All features are reproducible from municipality-level variables and can be computed in real-time in the web application.

In der App angezeigte Bezeichnungen (Deutsch):

• municipality_area_proxy → Gemeindegröße
• emp_per_resident → Arbeitsplatzquote
• foreigner_count_est → Ausländerpopulation

---

Final Selected Features

Feature Set for Final Model:

• rooms – number of apartment rooms
• area – living area in m²
• pop – municipality population
• pop_dens – population density (per km²)
• frg_pct – percentage of foreign residents
• emp – number of employees in municipality
• tax_income – taxable income per capita
• municipality_area_proxy – proxy for geographic size
• emp_per_resident – economic activity indicator
• foreigner_count_est – estimated foreigner count

---

Reason for Selection

Final model: RandomForestRegressor (tuned from iteration 2)
Justification:

• Highest cross-validated $R^2$ across all iterations (0.5509)
• Lowest generalization gap (CV Std = 0.1060 vs baseline 0.1195)
• Feature engineering improves predictive power by +0.0331 in $R^2$
• Tuned hyperparameters reduce overfitting (max_depth=12, min_samples_split=5)
• RMSE of CHF 674.54 acceptable for price range 750–8000

---

Preprocessing Steps (Iteration 1 → 2)

Data Cleaning

1. Load original dataset (apartments in canton Zurich)
2. Remove rows with missing values (dropna())
3. Remove duplicate rows (drop_duplicates())
4. Filter unrealistic prices: keep 750 ≤ price ≤ 8000 CHF
5. Filter invalid structures: keep rooms > 0 and area > 0

Feature Engineering (Iteration 2)

1. Compute municipality_area_proxy from pop and pop_dens
2. Compute emp_per_resident from emp and pop
3. Compute foreigner_count_est from pop and frg_pct
4. Combine with baseline features for final training

Evaluation Method

• 5-fold cross-validation
• Metrics: $R^2$, RMSE, MAE
• No separate validation set (full data used with CV)

---

Metric Definition

$R^2$ (Coefficient of Determination):
Proportion of variance in price explained by features. Range: [0, 1]. Higher is better.

RMSE (Root Mean Squared Error):
Square root of average squared prediction error. Units: CHF. Lower is better.

MAE (Mean Absolute Error):
Average absolute prediction error. Units: CHF. Lower is better.

---

Application & Deployment

• App Framework: Gradio
• App File: app.py
• Saved Model: models/apartment_price_model.pkl
• Deployment Platform: Hugging Face Spaces (URL to be updated)

How to Run Locally

cd Projekt1
uv run python app.py

---

Submission Checklist (Mandatory)

• Trained regression model available (models/apartment_price_model.pkl)
• New feature(s) added (iteration 2 feature engineering)
• Working web application (app.py)
• Documented iterative modeling process (2 iterations, tables + metrics)
• Completed README
• README uploaded to Hugging Face repository
• Public application link inserted above

---

Notes

• Baseline R² (0.5446) is competitive for real estate price prediction
• Feature engineering provides modest +0.0331 improvement in $R^2$
• Standard deviation drop (0.1195 → 0.1060) indicates more stable predictions
• Model saved and ready for deployment on Hugging Face Spaces