Neural Network Regression vs Classification

Comparing four feed-forward neural network architectures on regression and classification tasks using OpenML datasets.

Overview

This project trains and compares four feed-forward neural network configurations on two distinct tasks. Task 1 is a regression problem using OpenML dataset 216 (soccer goals); Task 2 is a multi-class classification problem using OpenML dataset 44 (iris). Each task trains all four architectures—varying the number of hidden layers, units per layer, and activation functions—for 200 epochs. Training and validation loss/accuracy curves are saved as PNGs, and a summary performance table is printed for each task.

Requirements

• Python 3.8+
• tensorflow >= 2.12
• scikit-learn >= 1.0
• matplotlib >= 3.5
• numpy >= 1.21

Installation

pip install -r requirements.txt

Project Structure

neural-network-regression-classification/
├── train_neural_nets.py    # Neural network training and comparison script
├── requirements.txt        # Python dependencies
├── .gitignore
└── README.md

Usage

python train_neural_nets.py

Steps performed:

1. Downloads OpenML datasets 216 (regression) and 44 (classification).
2. Task 1: trains 4 neural network configurations on the regression dataset for 200 epochs each; saves 4 MSE plots (regression_*.png).
3. Task 2: trains 4 neural network configurations on the classification dataset for 200 epochs each; saves 4 accuracy plots (classification_*.png).
4. Prints a performance table for each task showing best validation metric per configuration.

Results

Eight PNG plots are produced (4 regression, 4 classification):

Pattern	Description
regression_layers{N}_units{U}_{act}.png	Training vs validation MSE over 200 epochs
classification_layers{N}_units{U}_{act}.png	Training vs validation accuracy over 200 epochs

Two performance summary tables are printed to stdout.

Author

Biswajeet Sahoo

License

MIT License