Predictors Projects Collection 🤖

A comprehensive collection of machine learning and deep learning projects focused on predictive modeling across different domains: computer vision, medical diagnosis, insurance analytics, and multi-class classification.

📋 Table of Contents

• Overview
• Projects
  • California Housing Price Prediction
  • Medical Insurance Cost Predictor
  • Transfer Learning ResNet18
  • CIFAR-10 Neural Network
  • Diabetes Predictor
  • Flower Classification
• Technologies Used
• Getting Started
• Project Comparison
• Installation
• Usage
• Results Summary
• Repository Structure
• Contributing

🌟 Overview

This repository contains six distinct machine learning projects that demonstrate various approaches to predictive modeling:

1. Real Estate Prediction: Housing price prediction using advanced regression techniques
2. Insurance Analytics: Medical insurance cost prediction with feature engineering
3. Transfer Learning: Advanced computer vision with pre-trained ResNet18
4. Computer Vision: Image classification using Convolutional Neural Networks
5. Medical Prediction: Healthcare analytics with traditional ML algorithms
6. Multi-class Classification: Large-scale flower species recognition

Each project showcases different aspects of machine learning pipeline development, from data preprocessing to model evaluation, covering both traditional ML and deep learning approaches.

🚀 Projects

🏠 California Housing Price Prediction

Domain: Real Estate Analytics | Type: Regression | Framework: Scikit-learn + XGBoost

A comprehensive machine learning pipeline for predicting median house values in California districts using the famous California Housing dataset from scikit-learn.

Key Features:

• Dataset: 20,640 housing districts with 8 numerical features
• Best Model: XGBoost with 84.4% R² score - Excellent performance!
• Pipeline: Complete regression workflow with advanced preprocessing
• Evaluation: Multiple metrics (R², RMSE, MAE) with cross-validation
• Optimization: RandomizedSearchCV for hyperparameter tuning

Technical Highlights:

• Advanced Preprocessing: RobustScaler for outlier handling
• Model Comparison: Linear, Ridge, Lasso, Random Forest, XGBoost
• Cross-Validation: 10-fold CV for robust model selection
• Business Application: Suitable for Automated Valuation Models (AVM)

Models Performance:

Model	Cross-Validation R²	Test R²	RMSE	MAE
Linear Regression	~0.60	-	-	-
Ridge Regression	~0.60	-	-	-
Lasso Regression	~0.60	-	-	-
Random Forest	~0.80	-	-	-
XGBoost	~0.84	0.844	0.204	0.291

Business Value:

• 84.4% variance explained = Highly reliable price estimates
• Low prediction errors suitable for real estate valuation
• Scalable framework for property portfolio analysis

📁 Location: ./Machine Learning/California Housing/
📓 Main File: CaliforniaHousing.ipynb

---

🏥 Medical Insurance Cost Predictor

Domain: Insurance Analytics | Type: Regression | Framework: Scikit-learn + XGBoost

A comprehensive machine learning pipeline for predicting medical insurance costs based on personal characteristics and lifestyle factors using advanced feature engineering and hyperparameter optimization.

Key Features:

• Dataset: 1,338 insurance records with 7 features
• Best Model: XGBoost with 88.4% R² score - Outstanding performance!
• Pipeline: Complete regression workflow with categorical encoding
• Evaluation: Cross-validation with business-focused metrics
• Optimization: RandomizedSearchCV for hyperparameter tuning

Technical Highlights:

• Advanced Preprocessing: LabelEncoder + One-Hot encoding for categoricals
• Model Comparison: Linear, Ridge, Lasso, Random Forest, XGBoost
• Cross-Validation: 10-fold CV for robust model validation
• Business Application: Insurance premium calculation and risk assessment

Models Performance:

Model	Cross-Validation R²	Test R²	Notes
Linear Regression	~0.75	-	Baseline
Ridge Regression	~0.75	-	Regularized
Lasso Regression	~0.75	-	Feature Selection
Random Forest	~0.80	-	Ensemble
XGBoost	~0.87	0.884	Best Performance

Business Value:

• 88.4% variance explained = Extremely reliable cost estimates
• Feature insights for risk factor identification
• Scalable framework for insurance portfolio pricing

📁 Location: ./Machine Learning/Medical Insurance/
📓 Main File: Medical Insurance.ipynb

---

🔄 Transfer Learning ResNet18

Domain: Computer Vision | Type: Binary Classification | Framework: PyTorch + Transfer Learning

An advanced transfer learning implementation using pre-trained ResNet18 for cats vs dogs classification, demonstrating state-of-the-art computer vision techniques with multiple fine-tuning strategies.

Key Features:

• Dataset: 8,000 training + 2,023 test images (cats vs dogs)
• Best Model: Fine-tuned ResNet18 with 98.5% test accuracy - Outstanding performance!
• Architecture: Pre-trained ResNet18 with modified classifier head
• Strategies: Feature extraction, full fine-tuning, partial fine-tuning
• Optimization: Multiple learning rates for backbone vs head layers

Technical Highlights:

• Transfer Learning: ImageNet pre-trained ResNet18 backbone
• Advanced Data Augmentation: RandomResizedCrop, RandomHorizontalFlip
• Multiple Training Strategies:
  • Feature extraction (frozen backbone): 98.37% accuracy
  • Full fine-tuning: 98.56% accuracy
  • Partial fine-tuning (layer4 + fc): 98.32% accuracy
• Differential Learning Rates: Lower LR for backbone, higher for head
• Model Variants: 7 different saved models with various configurations

Training Strategies Performance:

Strategy	Test Accuracy	Notes
Feature Extraction	98.37%	Only FC layer trained
Full Fine-tuning	98.56%	Best Performance
Partial Fine-tuning	98.32%	Layer4 + FC trained
No Augmentation	~98.3%	Baseline comparison
No Warmup	~98.2%	Training strategy comparison

Model Artifacts:

• resnet18_finetuned.pth: Best full fine-tuning model
• resnet18_finetuned_warmup.pth: With warmup strategy
• resnet18_finetuned_noaug.pth: Without data augmentation
• resnet18_partial_finetune.pth: Partial fine-tuning variants

Business Value:

• 98.5% accuracy = Production-ready computer vision model
• Transfer learning efficiency = Faster training with less data
• Multiple strategies for different computational budgets
• Scalable framework for binary image classification tasks

📁 Location: ./Deep Learning/Transfer Learning ResNet18/
📓 Main File: Transfer Learning.ipynb

---

🖼️ CIFAR-10 Neural Network Classification

Domain: Computer Vision | Type: Multi-class Classification | Framework: PyTorch

An improved Convolutional Neural Network implementation for classifying images from the CIFAR-10 dataset into 10 distinct categories (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck) with enhanced architecture and training strategies.

Key Features:

• Dataset: 60,000 32×32 color images in 10 classes
• Architecture: Enhanced CNN with BatchNormalization and Dropout
• Performance: 78.51% accuracy on test set - Significant improvement!
• Training: 30 epochs with optimized hyperparameters
• Output: Saved model weights (cifar_net.pth)

Technical Highlights:

• Improved Architecture: 3-layer CNN with BatchNorm and Dropout
• Enhanced Preprocessing: Advanced data augmentation with larger batch size (32)
• GPU Acceleration: CUDA support for faster training
• Real-time Loss Monitoring: Training progress tracking
• Comprehensive Evaluation: Per-class accuracy analysis

Architecture Details:

• Conv Layers: 32→64→128 channels with 3×3 kernels and padding
• Batch Normalization: After each convolutional layer
• Dropout: 0.5 regularization before final classification
• Fully Connected: 512 hidden units for classification head

Performance Improvements:

• Previous: ~54% accuracy (basic CNN)
• Current: 78.51% accuracy (+24.5% improvement!)
• Training: 30 epochs vs. previous 20 epochs
• Batch Size: Increased from 4 to 32 for better convergence

📁 Location: ./CIFAR10/
📓 Main File: CIFAR10_neural_network.ipynb

---

🏥 Diabetes Predictor

Domain: Healthcare Analytics | Type: Binary Classification | Framework: Scikit-learn + XGBoost

A comprehensive machine learning pipeline for predicting diabetes risk using medical and demographic data from the Pima Indians Diabetes Database.

Key Features:

• Dataset: 768 patient records with 8 medical features
• Best Model: XGBoost with 76% recall, 72% F1-score
• Pipeline: End-to-end ML workflow with preprocessing
• Clinical Focus: High recall to minimize missed diagnoses
• Class Imbalance: Handled with SMOTE and scale_pos_weight

Technical Highlights:

• Advanced Preprocessing: Missing value imputation, robust scaling
• Feature Engineering: Glucose/Insulin ratio creation
• Model Comparison: Logistic Regression variants vs. XGBoost
• Medical Validation: Clinically relevant evaluation metrics
• Comprehensive EDA: Detailed data exploration and visualization

Models Trained:

Model	Recall	F1-Score	AUC	Notes
Logistic Regression	0.55	0.65	0.82	Baseline
Polynomial Features	0.60	0.68	0.83	Non-linear
XGBoost	0.76	0.72	0.85	Best

📁 Location: ./DiabetPredictor/
📓 Main File: DiabetPredictor.ipynb

---

🌸 Flower Classification with CNN

Domain: Computer Vision | Type: Multi-class Classification (102 classes) | Framework: PyTorch

A deep learning project for classifying flower images into 102 different species using the Oxford 102 Category Flower Dataset.

Key Features:

• Dataset: 8,189 high-resolution flower images
• Classes: 102 different flower species
• Architecture: 3-layer CNN with ~51.5M parameters
• Input: 224×224×3 RGB images
• Preprocessing: Resizing, normalization, tensor conversion

Technical Highlights:

• Custom Dataset Class: Handles MATLAB label files
• Large-scale Classification: 102-way classification problem
• GPU Optimization: CUDA acceleration for training
• Data Loading: Efficient batch processing
• Pre-defined Splits: Uses official train/validation/test splits

Model Architecture:

Input (224×224×3) → Conv+ReLU+MaxPool → Conv+ReLU+MaxPool → 
Conv+ReLU+MaxPool → Flatten → FC(512) → FC(102 classes)

📁 Location: ./Flower Classification/
📓 Main File: Flowers.ipynb

🛠️ Technologies Used

Machine Learning Frameworks:

• PyTorch: Deep learning models (Transfer Learning, CIFAR-10, Flower Classification)
• Scikit-learn: Traditional ML algorithms (Diabetes Predictor)
• XGBoost: Gradient boosting for tabular data
• Transfer Learning: Pre-trained model fine-tuning

Data Processing:

• Pandas: Data manipulation and analysis
• NumPy: Numerical computing
• SciPy: Scientific computing (MATLAB file handling)
• imbalanced-learn: Class imbalance handling (SMOTE)

Visualization:

• Matplotlib: Plotting and visualization
• Seaborn: Statistical data visualization

Development Environment:

• Jupyter Notebook: Interactive development
• Python 3.8+: Core programming language
• CUDA: GPU acceleration (optional)

🚀 Getting Started

Prerequisites

• Python 3.8 or higher
• pip package manager
• Jupyter Notebook
• GPU with CUDA support (recommended for deep learning projects)

Quick Start

1. Clone the repository

   git clone https://github.com/doroteaMonaco/Predictors-Projects.git
   cd Predictors-Projects

2. Choose a project and navigate to its directory

   cd CIFAR10                    # For CIFAR-10 classification
   cd DiabetPredictor           # For diabetes prediction
   cd "Flower Classification"   # For flower classification

3. Open the Jupyter notebook

   jupyter notebook

📊 Project Comparison

Aspect	California Housing	Medical Insurance	Transfer Learning	CIFAR-10	Diabetes Predictor	Flower Classification
Domain	Real Estate	Insurance Analytics	Computer Vision	Computer Vision	Healthcare	Computer Vision
Data Type	Tabular	Tabular	Images (224×224)	Images (32×32)	Tabular	Images (224×224)
Problem Type	Regression	Regression	Binary Classification	Classification	Binary Classification	Multi-class (102)
Samples	20,640	1,338	10,023	60,000	768	8,189
Algorithm	XGBoost	XGBoost	ResNet18 Transfer	CNN	XGBoost	CNN
Framework	Scikit-learn	Scikit-learn	PyTorch	PyTorch	Scikit-learn	PyTorch
Performance	84.4% R²	88.4% R²	98.5% Accuracy	78.5% Accuracy	76% Recall	In Progress
Focus	Price Prediction	Cost Prediction	Transfer Learning	Multi-class	Medical/Recall	Large-scale
Complexity	High Pipeline	High Pipeline	Advanced CV	Medium	High Pipeline	High Architecture

💻 Installation

Core Dependencies

# Essential packages for all projects
pip install jupyter pandas numpy matplotlib seaborn

# For PyTorch projects (CIFAR-10, Flower Classification)
pip install torch torchvision torchaudio

# For traditional ML (Diabetes Predictor)
pip install scikit-learn xgboost imbalanced-learn

# For flower classification MATLAB files
pip install scipy

# For visualization
pip install pillow

GPU Support (Optional but Recommended)

# For CUDA-enabled PyTorch (check CUDA version first)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Verify Installation

import torch
import sklearn
import xgboost
print("PyTorch version:", torch.__version__)
print("CUDA available:", torch.cuda.is_available())
print("Scikit-learn version:", sklearn.__version__)

📈 Usage

Running Individual Projects

California Housing Price Prediction:

cd "Machine Learning/California Housing"
jupyter notebook CaliforniaHousing.ipynb
# Uses built-in scikit-learn dataset

Medical Insurance Cost Predictor:

cd "Machine Learning/Medical Insurance"
jupyter notebook "Medical Insurance.ipynb"
# Uses included insurance.csv dataset

Transfer Learning ResNet18:

cd "Deep Learning/Transfer Learning ResNet18"
jupyter notebook "Transfer Learning.ipynb"
# Uses cats vs dogs dataset in archive/ folder

CIFAR-10 Classification:

cd "Deep Learning/CIFAR10"
jupyter notebook CIFAR10_neural_network.ipynb
# Dataset downloads automatically

Diabetes Predictor:

cd "Machine Learning/Diabet Predictor"
jupyter notebook DiabetPredictor.ipynb
# Uses included diabetes.csv dataset

Flower Classification:

cd "Deep Learning/Flower Classification"
jupyter notebook Flowers.ipynb
# Requires manual dataset setup (see project README)

Training Tips:

• GPU Recommended: Deep learning projects benefit significantly from GPU acceleration
• Memory Requirements: Ensure sufficient RAM (8GB+) and GPU memory (2GB+)
• Training Time: Varies from minutes (diabetes) to hours (deep learning)

📊 Results Summary

Performance Overview:

Project	Metric	Value	Significance
Transfer Learning	Test Accuracy	98.5%	Outstanding computer vision performance
California Housing	Test R²	84.4%	Excellent regression performance
Medical Insurance	Test R²	88.4%	Outstanding regression performance
CIFAR-10	Test Accuracy	78.5%	Strong CNN performance with improvements
Diabetes	Recall	76%	High medical relevance
Diabetes	F1-Score	72%	Balanced performance
Flower	Architecture	51.5M params	Large-scale classification

Key Achievements:

• Transfer Learning: Outstanding 98.5% accuracy for binary image classification
• Medical Insurance: Outstanding 88.4% R² score for insurance cost prediction
• California Housing: Excellent 84.4% R² score for regression prediction
• CIFAR-10: Strong 78.5% accuracy with enhanced CNN architecture (+24.5% improvement)
• Diabetes: Clinically relevant model with 76% recall (38% improvement over baseline)
• Flower: Complex 102-class classification with sophisticated preprocessing

📁 Repository Structure

Predictors-Projects/
├── README.md                           # This comprehensive overview
├── .gitignore                         # Git ignore patterns
│
├── Machine Learning/                   # Traditional ML & Regression Projects
│   ├── California Housing/            # Real Estate Price Prediction
│   │   ├── README.md                  # Project-specific documentation
│   │   └── CaliforniaHousing.ipynb    # Complete regression pipeline
│   │
│   ├── Medical Insurance/             # Insurance Cost Prediction
│   │   ├── README.md                  # Detailed project documentation
│   │   ├── Medical Insurance.ipynb    # Complete ML pipeline (88.4% R²)
│   │   └── archive/                   # Dataset directory
│   │       └── insurance.csv          # Medical insurance dataset
│   │
│   └── Diabet Predictor/              # Healthcare Analytics
│       ├── README.md                  # Detailed project documentation
│       ├── DiabetPredictor.ipynb      # Complete ML pipeline
│       └── data_lab9/                 # Dataset directory
│           └── diabetes.csv           # Pima Indians Diabetes Database
│
└── Deep Learning/                      # Neural Networks & Computer Vision
    ├── Transfer Learning ResNet18/     # Advanced Transfer Learning
    │   ├── README.md                  # Project documentation
    │   ├── Transfer Learning.ipynb    # Complete transfer learning pipeline
    │   ├── resnet18_finetuned.pth    # Best model (98.5% accuracy)
    │   ├── resnet18_finetuned_warmup.pth # Warmup strategy model
    │   ├── resnet18_finetuned_noaug.pth  # No augmentation model
    │   ├── resnet18_partial_finetune.pth # Partial fine-tuning model
    │   └── archive/                   # Cats vs Dogs dataset
    │       ├── training_set/          # Training images
    │       └── test_set/              # Test images
    │
    ├── CIFAR10/                       # Computer Vision - 10 classes
    │   ├── README.md                  # Project-specific documentation
    │   ├── CIFAR10_neural_network.ipynb # Main implementation
    │   ├── cifar_net.pth             # Saved model weights
    │   └── data/                      # CIFAR-10 dataset (auto-downloaded)
    │       └── cifar-10-batches-py/   # Extracted dataset files
    │
    ├── Fashion MINST/                 # Fashion item classification
    │   └── [Fashion MNIST files]      # Fashion classification project
    │
    └── Flower Classification/         # Computer Vision - 102 classes
        ├── README.md                  # Project documentation
        ├── Flowers.ipynb              # Deep learning implementation
        ├── imagelabels.mat           # Flower category labels
        ├── setid.mat                 # Train/test/validation splits
        └── 102flowers/               # Flower images dataset
            └── jpg/                  # 8,189 flower images

🎯 Learning Outcomes

This collection demonstrates:

Technical Skills:

• Deep Learning: CNN architecture design and training
• Traditional ML: Feature engineering and model selection
• Data Preprocessing: Handling missing values, scaling, imbalance
• Model Evaluation: Appropriate metrics for different domains
• Framework Proficiency: PyTorch and Scikit-learn expertise

Domain Knowledge:

• Computer Vision: Image classification challenges and solutions
• Healthcare Analytics: Medical data characteristics and evaluation priorities
• Multi-class Problems: Scaling to large number of categories

Best Practices:

• Reproducible Research: Comprehensive documentation and code organization
• Evaluation Focus: Domain-appropriate metrics (accuracy vs. recall)
• Pipeline Development: End-to-end ML workflow implementation

🤝 Contributing

Contributions are welcome! Areas for improvement:

Enhancement Opportunities:

• Transfer Learning: Pre-trained models for image classification
• Hyperparameter Tuning: Systematic optimization
• Cross-Validation: Robust evaluation strategies
• Ensemble Methods: Combining multiple models
• Deployment: Model serving and API development

How to Contribute:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

🙏 Acknowledgments

Datasets:

• CIFAR-10: Canadian Institute for Advanced Research
• Pima Indians Diabetes: UCI Machine Learning Repository
• Oxford 102 Flowers: Visual Geometry Group, University of Oxford

Frameworks:

• PyTorch Team: Deep learning framework
• Scikit-learn Contributors: Machine learning library
• XGBoost Developers: Gradient boosting framework

Educational Support:

• Politecnico di Torino: Academic context and guidance
• Open Source Community: Libraries and tools that made this possible

📞 Contact & Links

Author: Dorotea Monaco
Institution: Politecnico di Torino
GitHub: @doroteaMonaco
Repository: Predictors-Projects

Project-Specific Links:

• California Housing Price Prediction
• Medical Insurance Cost Predictor
• Transfer Learning ResNet18
• Diabetes Predictor
• CIFAR-10 Implementation
• Flower Classification

---

Built with 🧠 intelligence, ❤️ passion, and 🔬 scientific rigor

Last Updated: October 2025