Farah-frija/Tunisian-House-pricing-predictor

# 🏠 Tunisian Real Estate Price Prediction ## Overview This project predicts **house selling prices in Greater Tunis** using a **machine learning regression pipeline** enhanced by **LLM-assisted preprocessing** to handle incomplete and unstructured real estate listings. --- ## Team - Farah Frija - Emna Gharbi - Rima Zarrouki - Wala Ali --- ## Objectives - Unify real estate scraped data from multiple platforms - Clean and enrich incomplete listings - Engineer spatial features - Compare regression models - Build an interactive price prediction system --- ## Dataset - ~3710 instances scraped from real estate platforms - Sources: Tecnocasa, Mubawab, Tayara - Structured attributes + free-text descriptions ### Target - `prix` (selling price in TND) --- ## Data Preprocessing ### Classical Steps - Duplicate removal - Type normalization - Outlier handling - Basic missing value treatment | With outliers | Without outliers | |:---:|:---:| | | | *Figures showing data distribution before and after outliers removal.* ### LLM-Assisted Enrichment We used the **Groq API (`chatgpt-oss-120`)** to: - Complete missing attributes (rooms, amenities, standing) - Extract structured information from text descriptions - Reduce data loss due to incomplete listings > The LLM was used **only during preprocessing**, not for prediction. --- ## Geolocation Processing - Latitude and longitude were obtained using the **OpenStreetMap API** - Exact coordinates were available only for some sources - For others, geocoding was performed at the administrative (delegation) level --- ## Feature Engineering - Distance to Tunis center (`distance_center`) - Distance to the coast (`distance_coast`) - Spatial features from latitude / longitude - Price per square meter (`price_m2`) for data validation --- ## Modeling - **Type:** Supervised learning - **Task:** Regression Models evaluated: - Linear Regression - KNN - Tree-based models - **CatBoost Regressor** ### Best Model: CatBoo …