Boston House Price Prediction

What I Built#

I developed this project for a university machine learning course to practice taking a regression problem all the way from data analysis to a working web interface.

The application predicts residential property prices from 13 census attributes (like number of rooms, property tax rate, highway accessibility, and pupil-teacher ratios). I trained the models in Python using Scikit-Learn and deployed the inference endpoint using a lightweight Flask web application.

Click to view fullscreen

How It Works#

Web Browser Form
       │
       ▼
┌──────────────────────────────┐
│       Flask Web Server       │ (Parses input numbers into a NumPy array)
└──────────────┬───────────────┘
               │
               ▼
┌──────────────────────────────┐
│  Serialized Scikit Pipeline  │ (joblib.load('regmodel.pkl'))
│  ├── StandardScaler          │ (Applies pre-fit scaling)
│  └── Regression Estimator    │ (Linear / Ridge / Random Forest)
└──────────────┬───────────────┘
               │
               ▼
┌──────────────────────────────┐
│  HTML Template Prediction    │ (Renders estimated price back to user)
└──────────────────────────────┘

1. 01
   Data Exploration & Cleaning: I analyzed the 13 census features using Pandas, identifying right-skewed variables (like crime rate and highway accessibility) and checking feature correlations against median home prices.
2. 02
   Model Training & Comparison: I trained and evaluated multiple algorithms:

   • Linear Regression as a baseline.
   • Ridge Regression with L2 regularization to penalize large coefficients on correlated features.
   • Random Forest Regressor to capture non-linear relationships.
3. 03
   Pipeline Packaging: To ensure web requests don't fail due to missing or mismatched feature scaling, I bundled the StandardScaler and model weights together into a single .pkl file using joblib.
4. 04
   Flask Interface: A simple HTML form accepts user inputs and passes them to a Flask route (/predict_api and /predict), which returns the estimated property value.

Click to view fullscreen

What I Learned#

• Bundle preprocessing with your model: Saving the scaler alongside the model into a single pipeline file eliminates bugs where raw, unscaled numbers get passed to the estimator during web inference.
• Visualizing correlations saves time: Plotting correlation heatmaps early on helped me quickly see which features (like number of rooms and lower status population percentage) had the strongest impact on home prices.