🩺 Supervised Learning Model Comparison

Use Case: Diabetes Prediction | Classification Benchmark

---

📋 Table of Contents

• Overview
• Features
• Tech Stack
• Project Structure
• Dataset
• Models Compared
• Installation
• Usage
• Results
• Key Insights
• Acknowledgements

---

🎯 Overview

This project builds an end-to-end machine learning pipeline to predict diabetes onset using the Pima Indians Diabetes Dataset. Three supervised learning models were trained, evaluated, and compared across multiple train/test splits to identify the best-performing model for this healthcare classification task.

The workflow covers:

• Data loading and profiling
• Data cleaning and feature engineering
• Model training and evaluation
• Model comparison across different train/test splits
• Interactive prediction using a Streamlit web app

---

✨ Features

• Data Preprocessing Pipeline — Handles zero-value imputation in medical columns and applies feature scaling
• Multi-Model Evaluation — Compares Logistic Regression, Random Forest, and XGBoost
• Interactive Dashboard — Built with Streamlit for model comparison and prediction
• Feature Importance Analysis — Shows the most influential features for prediction
• ROC Curve & Confusion Matrix — Visual evaluation for classification performance
• Robustness Testing — Tests performance across multiple train/test splits

---

🛠️ Tech Stack

Category	Tools
Language	Python 3.10+
Core Libraries	pandas, numpy, scikit-learn, xgboost
Visualization	matplotlib, seaborn, plotly
App Framework	Streamlit
Version Control	Git & GitHub

---

📁 Project Structure

supervised_learning_project/
├── data_analysis.py              # Data profiling, preprocessing, model training, evaluation
├── app.py                        # Streamlit interactive application
├── requirements.txt              # Python dependencies
├── README.md                     # Project documentation
├── model_comparison_results.csv  # Exported metrics across all splits
├── scaler.pkl                    # Saved StandardScaler object
├── logistic_regression_model.pkl # Saved Logistic Regression model
├── random_forest_model.pkl       # Saved Random Forest model
├── xgboost_model.pkl             # Saved XGBoost model
└── diabetes_data.csv             # Processed dataset

---

📊 Dataset

• Source: Pima Indians Diabetes Dataset (UCI Machine Learning Repository)
• Records: 768 patient records
• Features: 8 input features
• Target: Outcome
  • 1 = Diabetes Positive
  • 0 = Diabetes Negative

Feature	Description
Pregnancies	Number of times pregnant
Glucose	Plasma glucose concentration
BloodPressure	Diastolic blood pressure (mm Hg)
SkinThickness	Triceps skin fold thickness (mm)
Insulin	2-Hour serum insulin (mu U/ml)
BMI	Body mass index (kg/m²)
DiabetesPedigreeFunction	Diabetes pedigree function
Age	Age in years

---

🤖 Models Compared

Model	Type	Best For
Logistic Regression	Linear Model	Interpretability, simple baseline, probability-based output
Random Forest	Bagging Ensemble	Strong default performance, feature importance, reduced overfitting
XGBoost	Boosting Ensemble	High predictive power, tabular datasets, model benchmarking

---

🚀 Installation

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/supervised_learning_project.git ## Replace YOUR_USERNAME with actual username
cd supervised_learning_project

2. Create and activate a virtual environment

python -m venv venv
source venv/bin/activate

For Windows:

venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

---

🎮 Usage

1. Run the data analysis pipeline

python data_analysis.py

This script will:

• Load the dataset
• Profile and clean the data
• Handle missing/invalid zero values
• Scale the features
• Train all three models
• Evaluate them on different train/test splits
• Save the trained models and results

2. Launch the Streamlit application

streamlit run app.py

The app includes:

• Data Overview — Dataset preview and distributions
• Model Comparison — Side-by-side metric comparison
• Detailed Metrics — Confusion matrix, ROC curve, and feature importance
• Prediction — Enter patient values and get a diabetes prediction

---

📈 Results

The models were evaluated across three train/test splits: 70/30, 80/20, and 90/10. Performance was measured using Accuracy, Precision, Recall, F1-Score, and ROC-AUC.

Performance Summary

Train/Test Split	Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC
70/30	Logistic Regression	0.745	0.672	0.531	0.593	0.836
70/30	Random Forest	0.745	0.672	0.531	0.593	0.827
70/30	XGBoost	0.745	0.645	0.605	0.624	0.794
80/20	Logistic Regression	0.701	0.587	0.500	0.540	0.813
80/20	Random Forest	0.740	0.652	0.556	0.600	0.817
80/20	XGBoost	0.753	0.660	0.611	0.635	0.804
90/10	Logistic Regression	0.727	0.615	0.593	0.604	0.834
90/10	Random Forest	0.779	0.692	0.667	0.679	0.874
90/10	XGBoost	0.805	0.714	0.741	0.727	0.868

Best Model by Split

Split	Best Model	Why It Performed Best
70/30	XGBoost	Highest F1-score (0.624) and better recall than the other two models
80/20	XGBoost	Best balance of accuracy, precision, recall, and F1-score
90/10	XGBoost	Highest accuracy (0.805) and highest F1-score (0.727)

Model Ranking

1. XGBoost — Best overall model across all train/test splits
2. Random Forest — Strong second-best model, especially on the 90/10 split
3. Logistic Regression — Good interpretable baseline, but lower predictive performance

Final Recommendation

For this diabetes prediction use case, XGBoost is the recommended model because it delivered the most consistent and strongest classification performance across all train/test splits.

• Choose XGBoost when predictive performance is the main priority
• Choose Random Forest when you want strong performance with easier interpretability
• Choose Logistic Regression when you need a simple and explainable baseline model

---

💡 Key Insights

• Feature engineering mattered: Zero values in medical columns such as Glucose, BloodPressure, Insulin, SkinThickness, and BMI were treated as invalid values and replaced using median imputation.
• F1-score was the key metric: Since this is a healthcare classification problem, F1-score is more useful than accuracy alone because it balances precision and recall.
• XGBoost was the most consistent model: It achieved the best F1-score across all three train/test splits.
• Random Forest was a strong alternative: It performed especially well on ROC-AUC and offers useful feature importance.
• Logistic Regression remained valuable as a baseline: It is easier to explain and useful in scenarios where interpretability matters more than raw performance.

---

👨‍💻 Author

Hariharaan Murugesan Salaja MSc Data Science & Analytics — University of Hertfordshire LinkedIn GitHub

---

🙏 Acknowledgements

• Dataset: UCI Machine Learning Repository — Pima Indians Diabetes
• Built as a supervised learning portfolio project
• Mentorship and guidance by Harshit Tripathi

---

Made with ❤️ by Hari