KeraProject - Neural Networks Experiments with Keras

Machine learning project implementing neural network experiments using Keras/TensorFlow for three different ML problems: breast cancer binary classification, car price regression, and sonar signal classification.

📋 Project Overview

This project includes three comprehensive machine learning analyses:

1. Breast Cancer Classification: Binary classification to distinguish malignant from benign tumors
2. Cars Regression: Regression to predict car prices
3. Sonar Classification: Binary classification to distinguish metal from rock using sonar signals

🗂️ Project Structure

kerasProject/
├── README.md                           # This file
├── requirements.txt                    # Python dependencies
├── ML_Theory_Summary.md               # Neural networks theory summary
├── ML2025_lab5_NeuralNets.pdf        # Lab documentation
│
├── breast_cancer_analysis.py          # Main script - Breast cancer
├── breastCancerModules/               # Breast cancer analysis modules
│   ├── __init__.py
│   ├── data_handler.py               # Data management and preprocessing
│   ├── models.py                     # Model architectures
│   ├── training.py                   # Training and experiments
│   └── visualization.py              # Plots and visualizations
│
├── cars_regression_analysis.py       # Main script - Car regression
├── carsModules/                       # Car regression modules
│   ├── data_handler.py
│   ├── model.py
│   ├── training.py
│   └── visualization.py
│
├── sonar_classification_analysis.py  # Main script - Sonar classification
├── sonarModules/                      # Sonar classification modules
│   ├── __init__.py
│   ├── data_handler.py
│   ├── model.py
│   ├── training.py
│   └── visualization.py
│
├── results/                          # Experiment results
└── venv/                            # Virtual environment

Installation and Setup

Prerequisites

• Python 3.8+
• pip

Installation

# Clone repository
git clone <repository-url>
cd kerasProject

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or
venv\Scripts\activate     # Windows

# Install dependencies
pip install -r requirements.txt

Main Dependencies

• TensorFlow 2.20.0
• Scikit-learn 1.7.1
• NumPy 2.2.6
• Pandas 2.2.3
• Matplotlib 3.10.6

📊 Implemented Experiments

1. Breast Cancer Classification

Dataset: Wisconsin Breast Cancer Dataset (569 samples, 30 features) Problem: Binary classification (Malignant/Benign)

Features:

• Architecture comparison: Basic (64→32→1) vs Funnel
• Experiments on epochs, batch_size, and architecture
• Metrics: accuracy, precision, recall, F1-score
• Complete visualizations with confusion matrix

Execution:

python breast_cancer_analysis.py

2. Cars Regression

Dataset: Car price prediction based on customer characteristics Problem: Regression (continuous value prediction)

Features:

• Age, Gender, Miles/day, Personal debt, Monthly income
• Architecture: 64→32→1 with linear activation
• Loss function: MSE (Mean Squared Error)
• Feature importance analysis

Execution:

python cars_regression_analysis.py

3. Sonar Classification

Dataset: 208 samples, 60 sonar features Problem: Binary classification (Metal/Rock)

Features:

• Base model and improved model with Dropout
• Early stopping to prevent overfitting
• Parameters: epochs=100, batch_size=5
• Prediction confidence analysis

Execution:

python sonar_classification_analysis.py

Architectures and Configurations

Activation Functions

• ReLU: Hidden layers (computational efficiency)
• Sigmoid: Binary classification output (0-1)
• Linear: Regression output (continuous values)

Optimizers

• Adam: Used in all experiments for adaptive convergence

Loss Functions

• Binary Crossentropy: Binary classification
• MSE: Regression

Regularization Techniques

• Validation Split: 80/20 for overfitting monitoring
• Early Stopping: Automatic halt when validation loss stops improving
• Dropout: In improved sonar model

📈 Results and Visualizations

Each experiment automatically generates:

1. Training History: Accuracy/loss plots over epochs
2. Classification Analysis: Confusion matrix, per-class metrics
3. Prediction Analysis: Confidence distribution, prediction examples
4. Comprehensive Reports: Complete performance analysis

Results are saved in the results/grafici/ folder.