MDFS is a library to assist in MultiDimensional Feature Selection (MDFS), i.e. feature selection that accounts for multidimensional interactions in the dataset. To learn more about MDFS, please visit the MDFS website.

This project is the implementation of the MDFS library for Python. Functionality-wise, it is aligned with the R version of the MDFS library, but the interface differs to make it more native to the Python ecosystem (i.e. pythonic) and to free it from early assumptions carried on for backward compatibility in R.

This software is released the same as the R MDFS library: under the GNU General Public License (GPL) v3.

The copyrights are held by Radosław Piliszek (the package maintainer and author), Abraham Kaczmarski (major contributor to the new interface), Krzysztof Mnich and Witold Rudnicki (authors of the MDFS method).

See the common changelog.

The library consists of a single package module: mdfs, which exports all the user-facing functionality.

The mdfs package module needs to be imported. Then, the main function to run is, aptly named, run. It accepts a numpy data matrix data and its corresponding decision, and returns a dictionary with the details of analysis, including the entry for relevant_variables which gives the indices of variables deemed relevant under chosen conditions.

The following list gives the translation between R functions and their Python counterparts.

• MDFS = run
• ComputeMaxInfoGains = compute_max_ig
• ComputeInterestingTuples = compute_tuples
• ComputePValue = fit_p_value
• Discretize = discretize
• GetRange = get_suggested_range
• GenContrastVariables = gen_contrast_variables

Function parameter names have been adjusted to avoid the dot (.), replacing it with an underscore (_).

There is no global seed in use. All functions depending on PRNG take a seed parameter.

Due to the way the Python-C interface is implemented in this library with numpy views, there is one quirk to be aware of. Functions returning a Structure subclass object do so without incurring a copy. Properties present on such objects return views, not copies. These views do not protect the result from being garbage collected (i.e., think of them as weak references to the underlying data). Thus, to avoid freed memory reads, keep the original structures around when using these views or copy data elsewhere as necessary. This quirk might be lifted in the future.