MKLpy

MKLpy is a framework for Multiple Kernel Learning (MKL) inspired by the scikit-learn project. The library encapsulates everything you need to run MKL algorithms, from the kernels computation to the final evaluation.

MKLpy contains:

• the implementation of MKL algorithms (EasyMKL, GRAM);
• kernel functions (polynomial, boolean kernels, and string kernels);
• various metrics and tools (kernel_alignment, radius, margin, spectral ratio...)

The main MKL algorithms implemented in this library are

Name	Short description	Status	Source
AverageMKL	Computes the simple average of base kernels	Available	-
EasyMKL	Fast and memory efficient margin-based combination	Available	[1]
GRAM	Radius/margin ratio optimization	Available	[2]
R-MKL	Radius/margin ratio optimization	Available	[3]
MEMO	Margin maximization and complexity minimization	Available	[4]
PWMK	Heuristic based on individual kernels performance	Avaible	[5]
FHeuristic	Heuristic based on kernels alignment	Available	[6]
CKA	Centered kernel alignment optimization in closed form	Available	[7]
SimpleMKL	Alternate margin maximization	Work in progress	[8]

See

The Tutorials section contains useful snippets of code to start with MKLpy.

---

Installation

MKLpy is available on PyPI:

pip install MKLpy

MKLpy requires pytorch, scikit-learn, and cvxopt installed.

---

Work in progress

MKLpy is under development! We are working to integrate several new features, including:

• further MKL algorithms, such as MEMO, and SimpleMKL;
• additional kernels for structured data (graphs, trees);
• efficient optimization

---

Known issues

• When dealing with normalized kernels, you need to (i) compute the complete kernel matrix (training + test examples) and to (ii) split the matrix into training and test matrices. Currently, you cannot directly compute the normalized kernel for training and test. This is not efficient and it will be fixed in the next releases.

• Some boolean kernels (DNF and CNF kernels) are currently disabled.

• The documentation for developers, containing directives and tools to develop novel algorithms and functionalities, is currently not available.

• We're fixing some issues related to SimpleMKL. The algorithm will be available in the next weeks.

---

Citing MKLpy

If you use MKLpy for a scientific purpose, please cite the following preprint.

@article{lauriola2020mklpy,
  title={MKLpy: a python-based framework for Multiple Kernel Learning},
  author={Lauriola, Ivano and Aiolli, Fabio},
  journal={arXiv preprint arXiv:2007.09982},
  year={2020}
}