📊 LG Aimers 5 - Display Defect Prediction

This repository contains the code and related materials from the LG Aimers competition, where the goal was to develop a machine learning model to predict defective displays in the manufacturing process. The competition focused on improving quality control by detecting potential defects early.

🏆 Ranked 62nd out of 1123 participants in the competition.

🔍 Project Overview

• Objective: Predict defective displays using data collected from the manufacturing process
• Technologies Used: Python, XGBoost, Optuna
• Pipeline: Data preprocessing → Feature engineering → Model training & tuning → Performance evaluation
• Optimization Strategies: Hyperparameter tuning with Optuna (1000 trials) and Seed Ensemble
• Evaluation Metric: F1-score
• Training Environment: Google Colab (T4 GPU)

⚠ Note: Due to competition regulations, the dataset used in this project cannot be shared.

📁 Directory Structure

📂 LG_Aimers_Display_Defect
│── preprocess.py    # Data preprocessing module  
│── optimize.py      # Hyperparameter tuning module (Optuna 1000 trials)  
│── train.py         # Model training module (XGBoost with Seed Ensemble)  
│── inference.py     # Inference module  
│── main.py          # Main script integrating all modules  
│── requirements.txt # Required dependencies  
│── README.md        # Project documentation  

🚀 Key Highlights

1. Data Preprocessing (preprocess.py)

   • Handling missing values and outliers
   • Feature scaling and selection

2. Hyperparameter Optimization (optimize.py)

   • Optuna (1000 trials) for tuning XGBoost hyperparameters

3. Model Training (train.py)

   • XGBoost as the primary model
   • Seed Ensemble for performance improvement
   • Trained using Google Colab's T4 GPU

4. Inference (inference.py)

   • Running predictions on new data

5. Main Execution (main.py)

   • Integrates all modules and runs the complete pipeline
   • ⚠ Users must set the PATH variable in main.py to specify their dataset location

📌 Usage Instructions

1. Clone the repository:

   git clone https://github.com/your_username/LG_Aimers_Display_Defect.git
   cd LG_Aimers_Display_Defect

2. Install dependencies:

   pip install -r requirements.txt

3. Set the dataset path in main.py:

   PATH = "your/dataset/path.csv"

4. Run the full pipeline:

   python main.py

🏆 Results & Insights

• Final Model: XGBoost with Seed Ensemble
• Best Hyperparameters: Optimized using Optuna (1000 trials)
• Performance Metric: Evaluated using F1-score
• Feature Importance Analysis

🔸 Handling Class Imbalance

Since the dataset was collected from real-world manufacturing data, defective samples were extremely rare, leading to a severe target imbalance problem.
To address this issue, significant effort was put into balancing the dataset, including using scale_pos_weight in XGBoost.

🔸 Why XGBoost?

Among various models tested, XGBoost was chosen because:

1. Tree-based models were the most practical under the given constraints.
2. Google Colab's T4 GPU was used for training, and XGBoost was well-suited for this environment.

📜 License

This project is licensed under the MIT License.