Diabetes Prediction Model Comparison

This project compares various machine learning models for predicting diabetes outcomes using a large-scale dataset with one lakh records and multiple features. It introduces a novel deep learning model, the Feature Utility, Grouping, and Adaptive fusion Network (FUGA-Net), for enhanced analysis.

Models Included

• Logistic Regression (optimized for large datasets)
• Decision Tree (with memory optimization)
• Random Forest (parallel processing enabled)
• KNN (optimized for large-scale data)
• Naive Bayes (memory-efficient implementation)
• SVM (with cache optimization)
• Ensemble (Voting Classifier with parallel processing)
• Feature Utility, Grouping, and Adaptive fusion Network (FUGA-Net) (novel deep learning model with feature fusion)

Project Structure

• data/: Contains the diabetes dataset (1 lakh records).
• outputs/plots/: Contains visualizations for model evaluation, including:
  • correlation/: Correlation heatmap (optimized computation).
  • distributions/: Feature distribution plots (with sampling).
  • confusion_matrices/: Confusion matrices for each model.
  • roc_curves/: ROC curves for each model and a combined plot (including FUGA-Net).
  • pr_curves/: Precision-Recall curves for each model (including FUGA-Net).
  • model_comparison/: Bar plots comparing model metrics.
• outputs/metrics/: Contains CSV files with model evaluation metrics (model_comparison_results.csv) and detailed metrics for FUGA-Net (fuga_net_detailed_metrics.txt).
• outputs/best_model/: Contains analysis files for key models, including fuga_net_analysis.txt and ensemble_analysis.txt.

Requirements

• Python 3.8+
• Core packages:
  • numpy
  • pandas
  • scikit-learn
  • matplotlib
  • seaborn
  • torch
• Performance optimization:
  • numba
  • pyarrow
  • fastparquet

Installation

1. Clone the repository.
2. Install the required packages:

   pip install -r requirements.txt

Usage

1. Run the FUGA-Net evaluation script to train and save its results:

   python evaluate_fuga_net.py

2. Run the main analysis script to train traditional models and generate comparisons:

   python diabetes_analysis.py

Key Features

• Memory-efficient data processing
• Parallel processing for faster computation
• Batch processing for large datasets
• Optimized model parameters
• Efficient visualization techniques
• Comprehensive model evaluation
• Novel Feature Utility, Grouping, and Adaptive fusion Network (FUGA-Net) implementation

Results

The evaluation results for all models are compiled in outputs/metrics/model_comparison_results.csv. Detailed metrics for FUGA-Net are in outputs/metrics/fuga_net_detailed_metrics.txt. Visualizations, including combined ROC and individual PR curves, are available in the outputs/plots/ directory. Analysis of key models is in the outputs/best_model/ folder.

Memory Optimization

The project implements several memory optimization techniques:

1. Efficient data types and downcasting
2. Batch processing for large predictions (especially in FUGA-Net)
3. Garbage collection and CUDA memory management
4. Sampling strategies for visualization
5. Parallel processing for faster computation

License

This project is licensed under the MIT License - see the LICENSE file for details.