Apppublic

tschool01/zurich-apartment-price

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes

App README

Apartment Price Prediction (Regression)

Task: Predict monthly rental prices for apartments in the canton of Zurich

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<br>- Drop duplicates<br>- Select 7 original features<br>- 5-fold CV	Linear Regression<br>Random Forest (`n_estimators=100`)	0.54 (LR)<br>0.49 (RF)	RMSE std: 144 (LR)<br>135 (RF)	Baseline	LR: ☐ Overfitting ☑ Good Fit<br>RF: ☑ Overfitting ☐ Good Fit
2	Improve generalisation	- Add `employment_rate` feature<br>- Replace LR with Gradient Boosting<br>- Tune RF hyperparameters<br>- 5-fold CV	Gradient Boosting (`n_estimators=200`, `lr=0.05`)<br>Tuned Random Forest (`n_estimators=200`, `max_depth=12`)	0.55 (GB)<br>0.57 (RF)	RMSE std: 128 (GB)<br>146 (RF)	+0.08 R² (RF vs iter 1 RF)<br>+0.03 R² (RF vs iter 1 LR)	☑ Overfitting ☐ Good Fit (reduced gap vs iter 1)

Notes

Metric: R² and RMSE (CHF) – 5-Fold Cross-Validation

Created Features:

—employment_rate = emp / pop – fraction of the municipal population that is employed (new feature, not in prior exercises)

Final Selected Features:

—rooms – number of rooms
—area – living area in m²
—pop – municipal population
—pop_dens – population density (per km²)
—frg_pct – percentage of foreign residents
—emp – number of employees in municipality
—tax_income – average taxable income
—employment_rate – emp / pop (engineered)

Reason for Selection: Selected based on availability in the BFS municipality dataset (enabling live lookup in the app), feature importance from the Random Forest, and cross-validation performance improvement when adding employment_rate. The new feature reduces the gap between train and test R², indicating better generalisation.

Preprocessing Steps

1.Load raw apartment listing CSV merged with BFS municipality socioeconomic data
2.Remove rows with missing values (dropna())
3.Remove duplicate rows (drop_duplicates())
4.Engineer new feature: employment_rate = emp / pop
5.Select final 8 features as input; price (CHF/month) as target
6.No feature scaling applied – tree-based models are scale-invariant

Final Model

RandomForestRegressor (tuned, Iteration 2)

Hyperparameter	Value
`n_estimators`	200
`max_depth`	12
`min_samples_leaf`	4
`random_state`	42

Selected based on lowest cross-validation RMSE in Iteration 2. Trained on the full dataset and saved as apartment_price_model.pkl.

Files

File	Description
`model_training.ipynb`	Full training notebook with both iterations
`app.py`	Gradio web application
`apartment_price_model.pkl`	Serialised final model + feature list
`bfs_municipality_and_tax_data.csv`	Municipality socioeconomic data (BFS)
`requirements.txt`	Python dependencies