🌆 California Real Estate AI: Property Valuation Engine

A production-ready machine learning repository demonstrating end-to-end data pipelines, advanced feature engineering, model serialization, and interactive deployment.

---

🌐 Live Interactive Application

Explore the deployed system live on Hugging Face Spaces:
👉 Launch the Live Property Estimator Space

🎥 User Interface & Demo

Bespoke dark-themed, glassmorphism-inspired GIS dashboard for real-time house price predictions.

---

🛠️ System Architecture & Data Pipeline

This project demonstrates software engineering best practices for data science, ensuring a clean separation of concerns by separating training pipelines (train.py) from serving applications (app.py), and pre-serializing models to optimize sub-second online inference.

graph TD
    A[California Census CSV Dataset] -->|1. Data Cleaning| B(Drop Missing Values)
    B -->|2. Feature Engineering| C[Log Transforms & Custom Ratios]
    C -->|3. Feature Scaling| D(StandardScaler Fit & Transform)
    D -->|4. Training| E[Multi-Variable Linear Regression]
    E -->|5. Serialization| F[Export model.pkl, scaler.pkl, columns.pkl]
    F -->|6. Production Serving| G[app.py Streamlit UI Engine]
    H[User Real-time Inputs] -->|7. Preprocess & Scale| G
    G -->|8. High-Speed Inference| I[Estimated Property Value Display]

Loading

---

📈 Engineering Highlights & Decision Log

1. Advanced Feature Engineering

• Logarithmic Smoothing: Census features like total_rooms, total_bedrooms, population, and households exhibit severe right-skewness. Applying $y = \log(x + 1)$ normalizes the distribution, ensuring the Linear Regression model handles outliers robustly.
• Domain-Specific Ratios:
  • bedroom_ratio ($\frac{\text{total bedrooms}}{\text{total rooms}}$): Helps capture the density of bedrooms, which highly correlates with house style and density.
  • household_rooms ($\frac{\text{total rooms}}{\text{households}}$): Captures the average room scale per housing unit.
• Categorical Categorization: One-hot encodes ocean_proximity to enable numerical model coefficients for geographical locations.

2. Model Serialization (Production Pattern)

Rather than training the model dynamically on every single API/page request, this system pre-serializes the trained LinearRegression model and StandardScaler state as .pkl binary files.

• Latency Advantage: Drops model load time to <1ms, compared to training on-the-fly which takes several seconds and burdens CPU resources.
• Reliability: Built with an elegant dual fallback structure: loads local .pkl files first, with an automated fallback to train on-the-fly from GitHub or local CSV if artifacts are missing.

---

📊 Model Performance & Interpretation

Evaluation Metrics

• R² Score (Coefficient of Determination): 0.6520 (Explains 65.2% of the variance in California property values)
• Mean Absolute Error (MAE): $49,852
• Root Mean Squared Error (RMSE): $68,340

Learned Coefficients (Feature Impact)

The model uncovers critical econometric insights based on standard census data:

Feature	Coefficient Sign	Economic Interpretation
Median Income	🟢 Positive (Strongest)	Higher neighborhood income is the most powerful driver of home valuation.
Inland Location	🔴 Negative (Strongest)	Properties situated inland suffer a massive valuation discount compared to coastal areas.
Near Ocean / Near Bay	🟢 Positive (Moderate)	Coastline proximity yields a substantial valuation premium.
Bedroom Ratio	🔴 Negative (Moderate)	Higher bedroom density (e.g. apartment blocks) indicates lower single-family premium pricing.

---

📁 Repository Directory Structure

House_Predicition_Values/
│
├── train.py                 # Offline Model training, evaluation & serialization pipeline
├── app.py                   # High-performance Streamlit UI & inference serving logic
├── requirements.txt         # Package dependencies with compatible pins
├── README.md                # Premium developer and recruiter documentation
├── housing.csv              # California Housing Census source dataset
│
├── models/                  # Serialized Production Artifacts
│   ├── model.pkl            # Serialized Scikit-learn Linear Regression model
│   ├── scaler.pkl           # Serialized StandardScaler state
│   └── columns.pkl          # Exported trained feature schema alignment
│
├── notebooks/               # Research & Prototyping
│   └── california_housing_analysis.ipynb  # Exploratory Data Analysis & training log
│
└── media/                   # Portfolio Assets
    ├── House hold Screenshot.png  # Interactive GIS Web App screenshot
    └── house_prediction_demo.mp4  # HD Video application walk-through

---

⚙️ Local Development Setup Guide

Follow these steps to run the training pipeline and launch the web interface locally:

1. Clone the Source Code

git clone https://github.com/bilalahmed251/House_Predicition_Values.git
cd House_Predicition_Values

2. Create and Activate Virtual Environment

# Windows
python -m venv venv
venv\Scripts\activate

# macOS / Linux
python3 -m venv venv
source venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

4. Execute the Training Pipeline

To retrain the model and regenerate the .pkl files inside models/:

python train.py

5. Launch the Streamlit Server

streamlit run app.py

The application will boot up at http://localhost:8501/ with live hot-reloading.

---

☁️ Serverless Cloud Deployment

This app is optimized for serverless hosting on Hugging Face Spaces (using the Streamlit SDK):

1. Create a new Streamlit Space on Hugging Face.
2. Push app.py, requirements.txt, housing.csv, and the pre-trained models/ folder.
3. Hugging Face automatically spins up the instance and serves the property valuation engine.

---

_{Engineered with 💙 for Data Science & Real Estate Analytics • MIT License 2026}