Apppublic

Danydarizzler/apartment-price-zurich

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes

App README

Model Iterations Documentation

Task: Apartment Price Prediction (Regression)

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>- Remove duplicates<br>- StandardScaler<br>- 5-fold CV	Linear Regression<br>Random Forest (n=100)	0.4739 (LR)<br>0.4648 (RF)	0.0980 (LR)<br>0.1275 (RF)	Baseline	☐ Overfitting ☑ Underfitting ☐ Good Fit
2	Improve with new feature & better models	- All steps from Iter 1<br>- New feature: `area_per_room`<br>- 5-fold CV	Ridge (alpha=10)<br>Gradient Boosting (n=200, depth=4, lr=0.05)	0.4786 (Ridge)<br>0.4977 (GB)	0.1099 (Ridge)<br>0.1005 (GB)	+0.0329 improvement	☐ Overfitting ☐ Underfitting ☑ Good Fit

New Feature

`area_per_room` = living area (m²) / number of rooms

—Represents the average room size of the apartment
—Not included in prior exercises (Week 1 and Week 2 used area and rooms separately)
—Motivation: a 3-room apartment with 90 m² (30 m²/room) is worth more than one with 60 m² (20 m²/room), even with the same room count

Preprocessing Steps

1.Load CSV data (original_apartment_data_analytics_hs24.csv)
2.Drop rows with missing values (dropna())
3.Remove duplicate rows (drop_duplicates())
4.Engineer new feature: area_per_room = area / rooms
5.Apply StandardScaler to all features (inside sklearn Pipeline)
6.Evaluate with 5-fold cross-validation (R² metric)

Models Used

Model	Iteration	Hyperparameters	Notes
Linear Regression	1	default	Baseline, prone to underfitting
Random Forest	1	nestimators=100, randomstate=42	Overfits without tuning
Ridge Regression	2	alpha=10.0	L2 regularization reduces overfitting
Gradient Boosting	2	nestimators=200, maxdepth=4, learningrate=0.05, randomstate=42	Best overall performer

Evaluation Method

—Metric: R² (coefficient of determination)
—Validation strategy: 5-fold cross-validation on the full dataset
—Higher R² = better fit (1.0 = perfect, 0.0 = predicts mean, <0 = worse than mean)

Final Selected Model

Gradient Boosting Regressor (Iteration 2)

Reason for selection:

—Highest CV R² across all models and iterations
—Handles non-linear relationships between features and price
—Lower overfitting compared to Random Forest thanks to shallow trees (max_depth=4) and slow learning rate (0.05)
—Benefits from the new area_per_room feature

Final features used (8 total):

—rooms – number of 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 – average taxable income in municipality
—area_per_room – new feature: average room size (area / rooms)

Application

The trained model is served via a Gradio web interface (app.py). Users select the number of rooms, living area, and municipality — the app returns the predicted monthly rent in CHF.