Apppublic

dubattim/zurich-apartment-price-predictor

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes

App README

Apartment Price Prediction - Model Iterations Documentation

Task: Apartment Rental Price Prediction (Regression)

Overview

This project predicts monthly rental prices for apartments in the Canton of Zurich using a Random Forest regressor. The model was developed iteratively over three iterations, progressively adding new features and tuning hyperparameters. The final model achieves an R² of 0.66 using 5-fold cross-validation.

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	- Outlier removal (price, area)<br>- Drop missing values<br>- 7 original features<br>- 5-fold CV	Random Forest (nestimators=100)<br>Ridge Regression (alpha=1.0)<br>Gradient Boosting (nestimators=100)	0.54 (Ridge)<br>0.51 (GB)<br>0.46 (RF)	0.09<br>0.06<br>0.05	Baseline	☑ Underfitting
2	Add new features + tuning	- Added distance_to_zurich (haversine distance to Zurich HB)<br>- Added area_per_room (derived feature)<br>- Hyperparameter tuning<br>- 5-fold CV	Tuned RF (nestimators=300, maxdepth=20)<br>Ridge (alpha=1.0)<br>Tuned GB (nestimators=300, maxdepth=5)<br>Lasso (alpha=1.0)	0.59 (RF)<br>0.53 (Lasso)<br>0.53 (Ridge)<br>0.52 (GB)	0.06<br>0.08<br>0.08<br>0.08	+0.13 (RF)	☐ Overfitting ☐ Underfitting ☑ Good Fit
3	Log-transform target	- Log1p transformation of target variable<br>- Enhanced features from Iter 2<br>- Increased n_estimators to 500<br>- 5-fold CV	RF (nestimators=500, maxdepth=20)<br>GB (nestimators=500, maxdepth=5, lr=0.05)	0.66 (RF)<br>0.61 (GB)	0.05<br>0.06	+0.07 (RF)	☐ Overfitting ☐ Underfitting ☑ Good Fit

Preprocessing Steps

1.Data loading: Loaded 804 apartment listings from enriched dataset (week 2), merged with BFS municipality demographic data
2.Missing value handling: Dropped rows missing essential features (rooms, area, price, population stats, coordinates)
3.Outlier removal: Removed apartments with price < 200 or > 15,000 CHF, area < 10 or > 500 m², rooms <= 0
4.Feature engineering:
5.distance_to_zurich (NEW): Haversine distance in km from apartment to Zurich main station (47.3769°N, 8.5417°E). Computed per listing, then averaged per municipality for the prediction app.
6.area_per_room: Living area divided by number of rooms
7.Target transformation (Iteration 3): Applied log1p to the rental price for training; predictions are converted back with expm1
8.String cleaning: Converted tax_income from formatted strings (e.g., "108'788") to float

Features Used (9 total)

Feature	Source	Description
rooms	User input	Number of rooms
area	User input	Living area in m²
pop	BFS data	Municipality population
pop_dens	BFS data	Population density
frg_pct	BFS data	Foreign resident percentage
emp	BFS data	Number of employed persons
tax_income	BFS data	Average taxable income
distancetozurich	Engineered (NEW)	Haversine distance to Zurich HB in km
areaperroom	Engineered	area / rooms

Feature Importances (Final Model)

Feature	Importance
area	0.4679
distancetozurich	0.2777
rooms	0.1148
areaperroom	0.0488
pop_dens	0.0246
tax_income	0.0242
pop	0.0233
frg_pct	0.0100
emp	0.0087

The newly engineered distance_to_zurich is the second most important feature (28%), confirming that proximity to the city center is a strong predictor of rental prices.

Evaluation Method

—Metric: R² (coefficient of determination)
—Validation: 5-fold cross-validation on the full dataset
—All models evaluated with the same CV splits for fair comparison

Final Selected Model

—Model: Random Forest Regressor
—Hyperparameters: nestimators=500, maxdepth=20, minsamplessplit=5, minsamplesleaf=2, random_state=42
—Target transformation: log1p (predictions converted back with expm1)
—CV R²: 0.6613 (±0.0501)
—Reason for selection: Highest cross-validation R² across all iterations and models. The log transformation improved performance by 0.07 R² over the non-log version.

Application

The Gradio web app (app.py) accepts:

—Number of rooms
—Living area (m²)
—Town (dropdown of 100+ Zurich-area municipalities)

It looks up municipality data from the enriched BFS dataset, computes area_per_room, and returns the predicted monthly rent in CHF.

Files

File	Description
`train_model.py`	Full training script with all 3 iterations
`app.py`	Gradio web application
`apartment_price_model.pkl`	Trained Random Forest model
`model_features.pkl`	Ordered feature list
`bfs_municipality_data_enriched.csv`	BFS data with distancetozurich
`bfs_municipality_and_tax_data.csv`	Original BFS municipality data
`requirements.txt`	Python dependencies