title	DefexHunter
emoji	🐛
colorFrom	blue
colorTo	indigo
sdk	docker
pinned	false
short_description	Predicts software defects from NASA JM1 static code metrics

🐛 DefexHunter ML

Production-ready REST API for software defect prediction using the NASA JM1 dataset.

---

What is this?

DefexHunter predicts whether a software module is defective or not based on static code metrics (Halstead complexity, cyclomatic complexity, line counts). It trains 7 ML models on the NASA JM1 dataset and serves predictions via a FastAPI REST API.

---

Live API

Base URL: https://asmaatef-defexhunter.hf.space
Docs:     https://asmaatef-defexhunter.hf.space/docs
Health:   https://asmaatef-defexhunter.hf.space/health
Models:   https://asmaatef-defexhunter.hf.space/models

---

Models

Model	Imbalance Strategy
XGBoost	scale_pos_weight + SMOTE
Random Forest	class_weight + SMOTE family / hybrid samplers
Decision Tree	class_weight + SMOTE
SVM	class_weight + SMOTE
Logistic Regression	class_weight + SMOTE
KNN	SMOTE
Naïve Bayes	SMOTE

All models are tuned via GridSearchCV (5-fold StratifiedKFold, macro F1 scoring).

---

Pipeline

Raw CSV → Clean → Stratified Split (70/30) → Feature Selection → StandardScaler → Resampling → GridSearchCV

---

Endpoints

Method	Endpoint	Description
GET	/health	Liveness check; lists loaded models
GET	/models	Returns training metrics for all models
POST	/predict	Single prediction
POST	/predict/batch	Up to 100 predictions in one call

---

Example Request

POST /predict

{
  "features": {
    "loc": 50,
    "ev(g)": 3,
    "iv(g)": 5,
    "n": 120,
    "l": 0.4,
    "d": 12.5,
    "i": 8.0,
    "t": 450.0,
    "lOCode": 40,
    "lOComment": 5,
    "lOBlank": 5,
    "locCodeAndComment": 2,
    "uniq_Op": 15,
    "uniq_Opnd": 20,
    "branchCount": 10
  },
  "model": "xgboost"
}

Example response:

{
  "prediction": 1,
  "label": "Defective",
  "probability": 0.87,
  "confidence": "High",
  "model_used": "xgboost"
}

---

Input Features

Feature	Description
loc	Lines of code
ev(g)	Essential cyclomatic complexity
iv(g)	Design complexity
n	Halstead program length
l	Halstead volume ratio
d	Halstead difficulty
i	Halstead intelligence
t	Halstead programming time
lOCode	Lines of code only
lOComment	Lines of comment
lOBlank	Blank lines
locCodeAndComment	Code + comment lines
uniq_Op	Unique operators
uniq_Opnd	Unique operands
branchCount	Branch count

---

Tech Stack

FastAPI · scikit-learn · XGBoost · imbalanced-learn · pandas · NumPy · Docker · Git LFS

Run Locally

# Train models first
python src/train.py data/jm1_csv.csv

# Run with Docker
docker-compose up --build

# Or directly
uvicorn main:app --host 0.0.0.0 --port 7860