🏗️ Steel Quality Prediction using Ensemble Learning

📝 Overview

This repository contains a high-performance machine learning pipeline designed to predict the quality score of steel products. In the manufacturing industry, predicting quality based on production parameters like temperature and cooling rate is crucial for reducing defects and optimizing processes.

The project was developed for a Kaggle competition hosted by Mohammad Saeid.

🚀 Key Features

• Advanced Feature Engineering: Extraction of time-based features (hour, weekday, weekend) and domain-specific interactions (e.g., Temperature/Time ratio).
• Automated Hyperparameter Tuning: Used Optuna to find the most efficient parameters for Gradient Boosting models.
• Model Stacking (Ensemble): Combined three powerful regressors (LightGBM, XGBoost, and CatBoost) using a RidgeCV meta-model to improve generalization and reduce RMSE.
• Robust Validation: Implemented a 5-fold Cross-Validation strategy to ensure the model's reliability on unseen data.

🛠️ Tech Stack

• Languages: Python
• Libraries: Pandas, NumPy, Scikit-learn
• Models: LightGBM, XGBoost, CatBoost
• Optimization: Optuna

📊 Pipeline Workflow

1. Data Preprocessing: Handling missing values and encoding categorical variables (Machine ID, Operator ID).
2. Feature Engineering: Creating physical interaction features like area_x_lum and temp_time_ratio.
3. Tuning: Running Optuna trials to minimize RMSE.
4. Ensembling: Training base models on 5 folds and blending their predictions using a Ridge regressor.

📈 Results

The final ensemble approach focuses on minimizing the Root Mean Squared Error (RMSE), providing a balanced prediction that captures complex patterns in the production line data.

📂 Project Structure

├── notebook.ipynb      # Main analysis and modeling code
├── submission.csv      # Final predictions
└── README.md           # Project documentation

👤 Author
Sheyda Asadi
⭐️ If you found this project helpful, please consider giving it a star!