# Support Feature Machines

**Authors:** Tomasz Maszczyk, Włodzisław Duch  
**arXiv:** [1901.09643](https://arxiv.org/abs/1901.09643)  
**Date:** 2019-01-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1901.09643v1/

## Key points

- Support Vector Machines (SVMs) with various kernels are widely used in machine learning for their attractive features.
- Interpreting solutions provided by SVMs can be challenging, especially when a single kernel type may not be suitable for all areas of the input space.
- Convergence issues can arise with certain kernels in SVMs.
- Standard quadratic programming solution for SVMs has time complexity of $O(m^3)$ and space complexity of $O(m^2)$ for m training patterns.
- Support Feature Machines (SFM) offer new and valuable features obtained from different kernels and vector transformations that can be applied in any machine learning algorithm.
- SFM is based on linear models in extended feature spaces, providing better control over the selection of support features compared to traditional SVMs.
- SFMs yield results comparable to kernel-based SVMs while addressing issues related to interpretation, scaling, and convergence.

## Summary

In their paper "Support Feature Machines," authors Tomasz Maszczyk and Włodzisław Duch discuss the significance of Support Vector Machines (SVMs) with various kernels in machine learning. These models have been widely used for many years across diverse applications, owing to their attractive features. However, interpreting the solutions provided by SVMs can be challenging. This is especially true when considering that a single kernel type may not be suitable for all areas of the input space. Additionally, convergence issues can arise with certain kernels. The standard quadratic programming solution for SVMs has time complexity of $O(m^3)$ and space complexity of $O(m^2)$ for m training patterns. are effective because they offer new and valuable features implicitly. These features, obtained from different kernels and vector transformations, can be directly applied in any machine learning algorithm. This facilitates the creation of multiresolution and heterogeneous data models. To address some of the limitations associated with traditional SVMs, Maszczyk and Duch propose . SFM is based on linear models in extended feature spaces, allowing for better control over the selection of support features. The authors demonstrate that SFMs yield results comparable to kernel-based SVMs while eliminating issues related to interpretation, scaling, and convergence. They showcase this through an analysis of benchmark datasets using linear discrimination, SVMs, decision trees, and nearest neighbor methods. The research presented in "Support Feature Machines" provides valuable insights into enhancing machine learning algorithms by leveraging linear models in extended feature spaces to improve performance and address key challenges encountered with traditional kernel-based approaches.