🐞 Software Bug Prediction using Semi-Supervised Learning

This project implements a semi-supervised learning approach to predict software bugs using the KC1 defect dataset from the PROMISE repository. It combines ensemble models with pseudo-labeling to leverage both labeled and unlabeled data, improving prediction performance under limited supervision.

---

🚀 Overview

• Developed a semi-supervised machine learning pipeline to detect defects in software modules.
• Applied SMOTE to handle class imbalance and StandardScaler for feature normalization.
• Used Random Forest feature importance for feature selection, achieving 88% accuracy on selected features.
• Combined Random Forest, XGBoost, SVM, and Logistic Regression in a soft-voting ensemble.
• Employed iterative pseudo-labeling with confidence thresholds (0.80–0.90) and early stopping based on validation F1-score stagnation.

---

🧱 System Architecture

The following pipeline summarizes the semi-supervised defect prediction architecture:

<📄 Data Input │ ▼ 🧪 Preprocessing (StandardScaler + SMOTE) │ ▼ 📊 Feature Selection (Random Forest Feature Importance) │ ▼ 🔁 Iterative Training Loop │ ├──► Train Ensemble Model │ ├── Random Forest │ ├── XGBoost │ ├── SVM │ └── Logistic Regression │ ├──► Predict Unlabeled Samples ├──► Select High-Confidence Pseudo Labels (Threshold: 0.80–0.90) └──► Augment Training Set │ ▼ ⏹ Early Stopping (Monitor Validation F1 Stagnation) │ ▼ ✅ Final Evaluation (Accuracy, Precision, Recall, F1-Score)

---

📊 Results

• Explored labeled data ratios from 10% to 90%.
• Achieved peak test performance:
  • Accuracy: 88.44%
  • F1-Score: 0.8451
  • Precision: 0.8491
  • Recall: 0.8411
• Demonstrated strong generalization using semi-supervised learning under label-scarce conditions.

---

🧰 Tools & Technologies

• Languages: Python
• Libraries:
  • scikit-learn (RandomForest, SVC, LogisticRegression, StandardScaler, metrics)
  • XGBoost
  • imbalanced-learn (SMOTE)
  • Pandas, NumPy
  • VotingClassifier

---

🗂 Dataset

• KC1 defect dataset from the PROMISE repository
• Preprocessing steps:
  • Handling missing values
  • Feature normalization
  • Binary classification based on defects field

---

🧠 Methodology

1. Preprocessing
   Normalize and balance data using StandardScaler and SMOTE.

2. Feature Selection
   Use Random Forest importance scores to select top predictors.

3. Model Training
   Train ensemble on labeled samples using soft-voting.

4. Pseudo-labeling
   Add confidently predicted unlabeled samples (≥ threshold) to the training set iteratively.

5. Early Stopping
   Monitor validation F1; stop if no improvement in 3 epochs.

6. Evaluation
   Assess final model using test set performance metrics.

---

📁 Files

• software_bug_prediction.py: Full implementation.
• kc1.csv: Input dataset.
• README.md: Documentation.

---

🙏 Acknowledgements

This project was conducted under the guidance of Ms. Vandana Bhattacharya as part of the Summer Training Program at Birla Institute of Technology, Mesra.
I would like to sincerely thank Ayush Anand (@Ayush-Dev09) and Ravi Kant for making this research project a fun and insightful experience. We are also thankful to the PROMISE repository for providing access to the KC1 software defect dataset used in this work.