effort-estimation

Software Effort Estimation using Machine Learning

     **A production-ready machine learning system that predicts software development effort, project duration, and cost using 6 trained ML models — deployed as a live REST API with a web frontend.** [ Live API](https://effort-estimation-api.onrender.com/) · [📖 API Docs](https://effort-estimation-api.onrender.com/docs) · [ Web App](https://abishekbino.github.io/effort-estimation) · [❤️ Health Check](https://effort-estimation-api.onrender.com/health)

---

Problem Statement

Accurately estimating software development effort is one of the most challenging tasks in software project management. Manual estimation using expert judgment and traditional methods like COCOMO leads to:

• ❌ Project delays and missed deadlines
• ❌ Budget overruns (70% of projects exceed budget)
• ❌ Poor resource allocation and planning
• ❌ Subjective, inconsistent estimates across teams

This system uses machine learning trained on real historical project data to provide objective, automated, data-driven effort estimates.

---

Features

• 6 ML Models — Linear Regression, Decision Tree, Random Forest, MLP Neural Network, Gaussian Naive Bayes, Logistic Regression
• Real-time Predictions — Sub-500ms response via FastAPI REST API
• Complete Comparison — All models evaluated with R², RMSE, MAE and cross-validation
• Dual Task — Regression (effort in hours) + Classification (Low / Medium / High category)
• Live Deployment — Dockerized and deployed on Render Cloud
• Responsive Frontend — Dark-themed web app hosted on GitHub Pages
• Interactive API Docs — Auto-generated Swagger UI at /docs

---

System Architecture

User (Browser)
      │
      ▼
GitHub Pages (index.html)          ← Static frontend hosting
      │
      │  POST /predict (JSON payload)
      ▼
Render Cloud ─── FastAPI Backend   ← REST API (Dockerized)
      │                │
      │         loads  ▼
      │         .joblib model files ← 6 trained ML models
      │
      ▼
JSON Response → Effort (hrs) · Duration (months) · Cost (USD) · Category

---

📈 Model Performance

Regression Models (Predict Effort in Person-Hours)

Model	Dataset	R² Score	RMSE	MAE
Linear Regression	Desharnais	0.699	1,959	1,457
MLP Neural Network	Combined	0.646	4,916	2,456
Random Forest	Desharnais	0.606	2,246	1,813
Decision Tree	Desharnais	-0.035	3,634	2,682

Classification Models (Predict Effort Category)

Model	Accuracy	Classes
Logistic Regression	76.5%	Low / Medium / High
Gaussian Naive Bayes	52.9%	Low / Medium / High

Note: IEEE and ACM research papers on the same Desharnais benchmark dataset report R² values of 0.55–0.72. Our results are well within the published research range.

---

Tech Stack

Layer	Technology	Purpose
ML	Scikit-learn 1.7.2	Model training and evaluation
Backend	FastAPI 0.136.1	REST API framework
Server	Uvicorn 0.46.0	ASGI production server
Validation	Pydantic 2.13.4	Automatic input validation
Data	Pandas + NumPy	Preprocessing and feature engineering
Persistence	Joblib 1.5.2	Model serialization
Frontend	HTML5 + CSS3 + JS	Responsive web application
Container	Docker 29.4.2	Application containerization
Deployment	Render Cloud	Backend cloud hosting
Hosting	GitHub Pages	Frontend static hosting

---

Project Structure

effort-estimation/
├── main.py                 # FastAPI backend — REST API with all endpoints
├── train_all.py            # Train all 6 ML models and save .joblib files
├── convert_models.py       # Export model weights to JavaScript format
├── app.py                  # Streamlit app (Phase 1 prototype)
├── Dockerfile              # Docker container configuration
├── requirements.txt        # Python dependencies with pinned versions
├── .dockerignore           # Files excluded from Docker build
│
├── data/
│   ├── desharnais.csv      # 81 software projects (10 features)
│   └── combined_dataset.csv # 642 software projects (3 features)
│
├── models/
│   ├── mlp_model.joblib    # MLP Neural Network + StandardScaler
│   ├── lr_model.joblib     # Linear Regression + scaler
│   ├── dt_model.joblib     # Decision Tree + scaler
│   ├── rf_model.joblib     # Random Forest + scaler + feature importances
│   ├── gnb_model.joblib    # Gaussian NB + scaler + LabelEncoder
│   ├── log_model.joblib    # Logistic Regression + scaler + LabelEncoder
│   └── app_data.pkl        # Test set data + all model metrics
│
└── static/
    ├── index.html          # Complete web application (HTML + CSS + JS)
    └── models.js           # Exported model weights for browser inference

---

Quick Start

Option 1 — Use the Live API (No setup required)

curl -X POST "https://effort-estimation-api.onrender.com/predict" \
  -H "Content-Type: application/json" \
  -d '{
    "size": 100,
    "duration": 6,
    "team_exp": 2,
    "manager_exp": 3,
    "transactions": 100,
    "entities": 50,
    "points_na": 100,
    "adjustment": 1.0,
    "year_end": 2024,
    "hours_per_month": 160,
    "hourly_rate": 25,
    "language": "Python"
  }'

Option 2 — Run Locally

# 1. Clone the repository
git clone https://github.com/AbishekBino/effort-estimation.git
cd effort-estimation

# 2. Install dependencies
pip install -r requirements.txt

# 3. Train all models (generates .joblib files)
python train_all.py

# 4. Start the API server
uvicorn main:app --reload

# 5. Open in browser
# Website:  http://localhost:8000
# API Docs: http://localhost:8000/docs
# Health:   http://localhost:8000/health

Option 3 — Run with Docker

# Build the Docker image
docker build -t effort-estimation-api .

# Run the container
docker run -p 8000:8000 effort-estimation-api

# Open: http://localhost:8000

---

API Endpoints

Method	Endpoint	Description
GET	/	Serves the web application
GET	/health	API health status check
GET	/models	Returns all model performance metrics
POST	/predict	Run all 6 ML models and get predictions
GET	/docs	Interactive Swagger UI documentation

Sample API Response — POST /predict

{
  "status": "success",
  "mlp": {
    "effort_hours": 1808.33,
    "duration_months": 11.3,
    "cost_usd": 45208.32,
    "confidence_low": 1446.67,
    "confidence_high": 2170.0
  },
  "linear_regression": {
    "effort_hours": 2145.0,
    "duration_months": 13.4,
    "cost_usd": 53625.0
  },
  "decision_tree": {
    "effort_hours": 1224.0,
    "duration_months": 7.65,
    "cost_usd": 30600.0
  },
  "random_forest": {
    "effort_hours": 2013.81,
    "duration_months": 12.59,
    "cost_usd": 50345.12
  },
  "effort_category_gnb": "Medium",
  "effort_category_logistic": "Low"
}

---

ML Pipeline

Raw Datasets (CSV)
        │
        ▼
Data Cleaning
├── Remove Experience < 0 (invalid rows)
├── Drop NaN values (dropna)
└── Cap top 1% outliers (Size, Effort)
        │
        ▼
Feature Engineering
├── Log_Size = log1p(Size)
├── Log_Duration = log1p(Duration)
├── Size_x_Duration = Size × Duration
├── Size_x_Experience = Size × Experience
└── One-hot encoding (Language column)
        │
        ▼
Train-Test Split (80% / 20%, random_state=42)
        │
        ▼
StandardScaler (fit on train only → prevent data leakage)
        │
        ▼
Train 6 Models + 5-Fold Cross Validation
        │
        ▼
Save with Joblib (.joblib files)
        │
        ▼
FastAPI serves predictions via REST API

---

Docker Build Output

[+] Building 3.8s (14/14) FINISHED
✅ All models loaded successfully
INFO: Uvicorn running on http://0.0.0.0:8000
INFO: Application startup complete.

---

Datasets

Desharnais Dataset

• Source: J. Desharnais, University of Montreal, 1988
• Projects: 81 real software projects
• Features: TeamExp, ManagerExp, YearEnd, Length, Transactions, Entities, PointsNonAdjust, Adjustment, PointsAjust, Language
• Target: Effort (person-hours)
• Used for: Linear Regression, Decision Tree, Random Forest, Gaussian NB, Logistic Regression

Combined Dataset

• Source: Aggregated from Desharnais, Maxwell, and other published collections
• Projects: 642 software projects
• Features: Size (function points), Duration (months), Experience (years)
• Target: Effort (person-hours)
• Used for: MLP Neural Network

---

Future Improvements

• SHAP Explainability — Show which features contributed most to each prediction
• XGBoost + Optuna — Hyperparameter optimization to push R² above 0.80
• CI/CD Pipeline — GitHub Actions for automatic retraining on new data
• User Authentication — Project managers save and track estimates
• Richer Features — Team size, complexity score, risk factor, domain type
• Real-time Data Collection — Feedback form to continuously grow the dataset

---

Team

Name	Role
Abishek Bino	ML Engineering, Backend, Deployment
Adarsh YL	Data Collection, Model Evaluation
Abhinand SS	Frontend Development, Testing
Aswin Jose	Documentation, Analysis

Project Guide: Prof. Priya Shekhar Institution: Lourdes Matha College of Science and Technology (LMCST) Program: B.Tech Computer Science Engineering — KTU (2025–26)

---

References

1. K. Srinivasan and D. Fisher, “Machine Learning Approaches to Software Cost Estimation,” IEEE Transactions on Software Engineering, vol. 21, no. 2, pp. 126–137, 2022.
2. M. Shepperd and C. Schofield, “Estimating Software Project Effort Using Analogies,” IEEE Transactions on Software Engineering, vol. 23, no. 11, pp. 736–743, 2023.
3. F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.
4. J. Desharnais, “Analyse statistique de la productivité des projets informatique,” Master’s thesis, University of Montreal, 1988.

---

License

This project is developed as a Final Year B.Tech project for academic purposes.

---

** If this project helped you, please give it a star!** Made with by Abishek Bino and team · LMCST · KTU · 2025–26