Kirazul/TunisiaRealEstate_ML

A complete end-to-end machine learning pipeline for predicting property prices across Tunisia. # Tunisia Real Estate Price Prediction System: A complete end-to-end machine learning pipeline for predicting property prices across Tunisia. The system combines data processing, geographic alignment, feature engineering, and gradient boosting to deliver accurate price predictions with full national coverage through intelligent fallback mechanisms. **Live Demo**: tunisiarealestate-ml.onrend… **Live Presentation**: tunisiarealestate-ml.onrend… --- ## Problem Statement 🎯 Real estate pricing in Tunisia varies dramatically by location. A 100 m2 apartment in Tunis costs fundamentally differently than the same property in Sfax or Kebili. This system addresses the challenge of predicting property prices across 264 delegations, where many regions lack sufficient training data, by combining direct modeling with hierarchical fallback benchmarks. --- ## Architecture Overview 🏗️ The system consists of three interconnected layers: | Layer | Description | |-------|------------| | **Data Pipeline** | Eight-stage transformation from raw listings to trained model | | **Prediction API** | FastAPI serving real-time predictions and model diagnostics | | **Geographic Atlas** | Interactive frontend displaying coverage, predictions, and fallback tiers | All components share common data artifacts, ensuring consistency between training and serving. --- ## Pipeline Stages 🔄 The pipeline transforms raw property listings through eight sequential stages: ### Stage 1: Dataset Discovery 🔍 Location: `pipeline/01_dataset_discovery.py` > Loads and profiles raw data sources. Identifies column mappings for price, surface, location, and property type across three source datasets. Produces discovery profiles documenting data quality, value distributions, and schema assessments. ### Stage 2: Dataset Cleaning 🧹 Location: `pipeline/02_dataset_cleaning.py` > Applies standardization rules across datasets. Handles encoding issues, trims whitespace, n …