Nested Dichotomy in Image Classification

This repository contains a series of Jupyter notebook experiments focused on multi-class image classification with a nested dichotomy strategy and a DecisionTreeClassifier base learner.

The project investigates how model performance changes depending on:

• feature extraction method,
• data balancing strategy,
• hyperparameter selection method,
• implementation details of nested dichotomy.

Project Goals

The notebooks answer four research questions:

1. exp1.ipynb – How different feature extraction methods affect the effectiveness of classification?
2. exp2.ipynb – How data balancing methods affect model quality?
3. exp3.ipynb – How parameter selection affects model quality?
4. exp4.ipynb – How nested dichotomy implementation affects model quality?

Methods Used

• Classifier: DecisionTreeClassifier (scikit-learn)
• Multi-class strategy: Nested Dichotomy
• Cross-validation: K-Fold (typically 5-fold)
• Feature extraction:
  • VGG16
  • InceptionV3
  • MobileNetV2
• Balancing methods (selected experiments):
  • SMOTE
  • RandomOverSampler
  • RandomUnderSampler
  • TomekLinks
  • SMOTETomek
• Hyperparameter optimization (selected experiments):
  • GridSearchCV
  • RandomizedSearchCV
  • BayesSearchCV
• Evaluation metrics:
  • Accuracy
  • Precision (weighted)
  • Recall (weighted)
  • F1-score (weighted)
  • Per-class accuracy

Repository Structure

.
├── exp1.ipynb
├── exp2.ipynb
├── exp3.ipynb
├── exp4.ipynb
├── requirements.txt
└── README.md

Dataset Layout

The notebooks expect image data in a local directory structure similar to:

data/
├── train/
│   ├── class_1/
│   ├── class_2/
│   └── ...
├── test/
│   ├── class_1/
│   ├── class_2/
│   └── ...
├── validation/
│   ├── class_1/
│   ├── class_2/
│   └── ...
└── augmented/
    ├── class_1/
    ├── class_2/
    └── ...

Each class folder should contain .jpg images (some notebooks also include .jpeg files from augmented/).

Setup

1. Create and activate a Python virtual environment.
2. Install dependencies:

pip install -r requirements.txt

3. Ensure the dataset is placed in the expected data/ subfolders.

Running Experiments

Open the notebooks in JupyterLab/VS Code and run cells sequentially:

• exp1.ipynb – feature extraction comparison
• exp2.ipynb – class balancing comparison
• exp3.ipynb – hyperparameter search comparison
• exp4.ipynb – nested dichotomy implementation variants

The notebooks generate printed metric reports and visualizations (e.g., bar plots, radar charts, confusion-related outputs).

Reproducibility Notes

• K-Fold splitting is configured with a fixed random seed (random_state=42) in the notebooks.
• Some model/training operations may still include non-deterministic behavior depending on library/hardware configuration.
• Results can vary if the dataset composition or augmentation content differs.

Requirements

All Python dependencies are listed in requirements.txt. Key libraries include:

• TensorFlow / Keras
• scikit-learn
• imbalanced-learn
• scikit-optimize
• NumPy, pandas, SciPy, matplotlib