🌧️ Project Overview

This project was developed for the Kaggle “Binary Prediction with a Rainfall Dataset” competition. The goal was to predict the occurrence of rainfall using a set of environmental features. My final solution achieved a public ROC AUC score of 99.95, placing it among the top scores in the competition.

🔍 Problem Scope

Generate predictions for the full test set with the probability of rain conditional to the available regressors (e.g. temperature, wind speed, cloud cover, and humidity).
60% of the test set is used only at the end of the competition, when the private score is computed, the remaining 40% is used to calculate the public score with each submission (it’s unknown which samples are part of the public score computation.)

🛠️ Technical Approach

✅ Feature Engineering

  • Normalized key predictors and engineered meaningful ratios and deltas.
  • Created visualizations to validate distributional consistency between training and test data.
  • Analyzed correlations with the target to prioritize high-signal features.

✅ Modeling

  • Trained two high-performing models:
    • Random Forest Classifier
    • XGBoost Classifier, using GPU acceleration via CUDA for faster tuning and inference
  • Used Optuna for Bayesian hyperparameter optimization with custom search spaces.

✅ Validation Strategy

  • Applied Grouped KFold and Random KFold to ensure the model generalized across known splits.
  • Carefully monitored overfitting using out-of-fold performance.

🚀 Deployment

To demonstrate the model’s predictions interactively and reproducibly, I:

  • Built a Streamlit app that allows exploration of predictions and feature importances.
  • Packaged the entire app using Docker to simplify reproducibility.
  • Hosted the application on a private cloud server accessible here:

👉 Live App

📊 Visual Preview

Screenshot of the Rainfall App

🧠 Lessons Learned

  • GPU acceleration can be very valuable even for structured data when doing intensive feature selection and hyperparameter optimization.
  • Cross validation is still one of the best tools in the ML field.
  • Binary classification with Random Forest and XGboost are still very powerful tools.

📁 Links

If you’d like help writing the missing “Lessons Learned” section or a custom visual banner, I can help with that too!