Author: Gintaras Koncevicius
Published on: ResearchGate | Blogspot
Least Similar Spheres (LSS) reimagines traditional Support Vector Machines (SVMs) by using geometric spheres instead of linear hyperplanes. This approach allows for more intuitive data classification in certain datasets, particularly bioassay datasets where traditional SVMs may struggle.
Awarded "Project of the Year" at ACE Showcase and published on ResearchGate.
- Spherical Classification Boundaries: Utilizes sphere radii for classification rather than support vectors, simplifying the visual interpretation of classification.
- Custom Distance Measurement: Multiple distance formulas implemented, allowing fine-tuning based on dataset requirements.
- Enhanced Data Handling: Incorporates slack variables, outlier removal, and cluster averaging to improve classification accuracy.
- Comparable or superior performance to traditional SVMs in specific datasets.
- Exact match with R-SVM predictions in certain bioassay dataset instances.
- Efficiently handles overlapping data points using customizable slack variables.
LSS replaces the traditional SVM approach with spheres, calculating classification boundaries based on the radius of each sphere relative to the dataset's features.
- Data Preprocessing: Outliers are removed, and data points are clustered for smoother classification.
- Sphere Generation: Spheres are generated using various distance measurement formulas.
- Classification: Based on sphere overlap and slack variable adjustments, data is classified with enhanced flexibility compared to traditional SVMs.
- Download: Clone or download the repository.
- Run: Execute
LSS.jarto launch the GUI and load sample data in CSV format. - Parameters: Set classification parameters (distance formula, slack variables) based on your dataset.
Here are some examples illustrating how the LSS algorithm visualizes classification boundaries:
Figure: Sphere-based boundaries for initial classification.
Figure: Outliers removed for cleaner data.
For a more in-depth look at the methodology and testing results:
- ResearchGate Publication: Complete paper with mathematical explanations and experimental outcomes.
- Blogspot Post: Non-technical overview and visual guide to the LSS approach.
- Interactive real-time visualization of classification boundaries.
- Additional distance measurement formulas for broader use cases.
- Command-line functionality for batch processing large datasets.
Interested in expanding the LSS project? Feel free to fork the repository or reach out for collaboration.
Thank you for exploring the Least Similar Spheres (LSS) project! For questions or collaboration opportunities, please contact me directly.