Iterative_Subset_Hyperblock_Search

Finding representative hyperblocks by iterating on subsets of the training data for repeating patterns.

Video of picking up completely randomly 1/3 subsets of the Fisher Iris training data iteratively with replacement: https://www.youtube.com/watch?v=HR7-bnE_b64

Working demo of growing 1/3 subset until HB purity violation:

How to use

1. Clone the repository:

   git clone https://github.com/yourusername/Iterative_Subset_Hyperblock_Search.git
   cd Iterative_Subset_Hyperblock_Search

2. Install the required dependencies:

   pip install pandas numpy matplotlib scikit-learn tabulate

3. Run the main script:

   python direct.py

   • You'll be prompted to choose between generated Gaussian test data or loading from a file
   • If you select option 1, the program will generate synthetic Gaussian data with two classes
   • If you select option 2, a file dialog will open to select your CSV dataset
   • The program will analyze the data and generate hyperblocks using three algorithms: IHyper, MHyper, and IMHyper
   • Results will be displayed in the console and visualized in parallel coordinates

4. Generating Gaussian test data:

   • When you select option 1, the program will generate synthetic data with two classes
   • By default, it creates 100 samples with 2 features
   • Class 0 follows a Gaussian distribution with mean at origin and unit variance
   • Class 1 follows a Gaussian distribution with mean at (3,3) and 1.5 variance
   • The generated data will be visualized before hyperblock analysis

5. For dataset splitting:

   python data_splitter.py

   • This utility will help you create multiple train-test splits for cross-validation
   • By default, it creates 10 splits with 1/3 training and 2/3 testing data
   • Stratified sampling is used to maintain class distribution

Dataset Preparation

Prepare your dataset with the following guidelines:

• The data should be in .CSV format
• One column should contain the class labels (by default, looks for a column named "class" case-insensitive or fallsback to using the last column)
• All other columns should be numeric features i.e. attributes

References

This project is working with algorithms available in:

[1] Kovalerchuk B, Hayes D. Discovering interpretable machine learning models in parallel coordinates. In 2021 25th International conference information visualisation (IV) 2021 Jul 5 (pp. 181-188). IEEE.
[2] Huber L, Kovalerchuk B, Recaido C. Visual knowledge discovery with general line coordinates. In Artificial Intelligence and Visualization: Advancing Visual Knowledge Discovery 2024 Apr 25 (pp. 159-202). Cham: Springer Nature Switzerland.
[3] Huber, Lincoln, "Human-Centered Machine Learning with Interpretable Visual Knowledge Discovery" (2024). All Master's Theses.

License

This project is licensed for free and commercial use under the MIT License. See the LICENSE file for details.