Project · Tabular · 2024 · team of four
Real-estate price prediction
Predicting French property prices from 40,000 incomplete listings, with nearest-neighbour imputation and a tuned XGBoost regressor.
Problem
Predict a property’s price from its listing (ENS Challenge Data no. 68): city, type, surfaces, rooms, energy
and greenhouse-gas ratings, orientation. The target is widely dispersed and the inputs are incomplete:
CITY alone has about 8,000 distinct values, and more than 18,000 listings have no energy rating.
Method
Ordinal ratings were label-encoded and nominal variables one-hot encoded. Missing surfaces and room counts were
filled with a K-nearest-neighbours imputer. Linear regression, KNN and a random forest were compared with
XGBoost, which was tuned with GridSearchCV (1,600 trees, depth 10, learning rate 0.04).
Results
XGBoost was the strongest model of the four. Adding listing photos as inputs did not improve the score.
Stack
Python · XGBoost · scikit-learn · pandas · Jupyter