fcmclt

Fuzzy clustering C (MEX) implementation for MATLAB (FCM, Gustafson-Kessel, clustering validity, fuzzy partition matrix extrapolation)

fcmclt

Fast fuzzy clustering C (MEX API) implementation for MATLAB (FCM, Gustafson-Kessel, clustering validity, extrapolation with presumed cluster centers)

Project status

The only planned changes at this time are to improve the documentation, i.e. this README file.

Description

This repository provides efficient C implementation of two types of Fuzzy c-Means Clustering algorithms and several related tools that can be used in clustering evaluation and validation. Once installed, the provided functions are directly callable from MATLAB and will run considerably faster compared to similar fuzzy clustering packages that are implemented purely in MATLAB language. To use these algorithms outside of MATLAB, one would need to modify the existing C code. Additional implementation details that may be useful are included in the implementation section below.

The following functionality is supported:

• Fuzzy c-Means Clustering (based on the work of James C. Bezdek and Enrique Ruspini.)
  • Euclidean, diagonal, and Mahalanobis distance metric
• Gustafson-Kessel fuzzy c-means clustering
• Initial fuzzy partition matrix generation
• 12 Clustering validity functionals:
  • partition coefficient, partition entropy, nonfuzzy index, minimum and mean hard tendencies, minimum and maximum relative fuzziness, the minimum nearest maximum membership cardinality, compactness and separation index, fuzzy hypervolume, average partition density, and the partition density of a resulting fuzzy partition
• Fuzzy partition matrix extrapolation that uses the presumed cluster centers
• Fuzzy scatter and covariance matrices calculation for fuzzy clusters
• Test code (in MATLAB) to verify the proper installation of software
• Detailed help text describing how to use the provided functions

The primary motivation for this work was to estimate membership functions of the Temporal Fuzzy Sets which span the Fuzzy Information Space. Nevertheless, the provided software package may be used for solving any problem that requires fuzzy c-means clustering, e.g. tracking of speech and noise power changes, echo path estimation, analysis, and characterization, etc. Fuzzy clustering has been applied within a wide range of engineering and research disciplines including geology, agriculture, social sciense, etc.

Installation

The installation may be done either by downloading the latest release which includes generated binaries for the Windows x64 MEX files (MATLAB 2022b used), or by cloaning or downloading the source code. The process of generating your own binaries from the source code is straighforward.

Prerequisites

To use this software one must have MATLAB installed and if you plan to generate the MEX binaries you must be able to run the mex command. This command is used for generating the accelerated versions of the three main algorithms from their C source code.

Steps to install

From the release

Once you verified you have MATLAB installed you can download the latest release. All the necessary files are contained within the fcmclt_M_m_p_w64.zip file. When you unpack the zip archive, the MATLAB scripts and the binaries will be contained in the ./src folder. You may optionally copy the contents of ./src directory to a folder where you want MATLAB to access it. As an example, let us assume that you copied all the files from ./src to D:\MyMatlab\fcmclt on a Windows system.

You may now proceed to testing your installation as described below.

Manual from the source with generating MEX binaries

Once you verified you have MATLAB installed and you can successfully run mex command to build MATLAB functions implemented in C language, you may proceed with the following steps:

• Clone or download this repository to the system where you have MATLAB installed
• Copy the contents of ./src directory to a folder where you want MATLAB to access it
  • As an example, let us assume you copied all the files from ./src to D:\MyMatlab\fcmclt on a Windows system.
• Start MATLAB and within MATLAB change directory to D:\MyMatlab\fcmclt
  • >> cd D:\MyMatlab\fcmclt
• Run the following three mex commands:

  mex extfpm.c
  mex fcmc.c
  mex gkfcmc.c

• You should see three new files created: extfpm.<mexext>, fcmc.<mexext>, gkfcmc.<mexext>, where <mexext> reflects the mex file extension on your system. For 64-bit Windows system, it should be mexw64.

Testing and finalizing your installation

• You can test your installation by running test routines that were provided in ./src folder and that you copied to D:\MyMatlab\fcmclt, e.g. try running >> testfcmc from that folder.
• Try all other test routines. They have prefix test or tst.
  • NOTE: Some may display a warning indicating the dataset may be too small for reliable clustering. That is normal and expected behavior for those examples.

If you want to make the fcmclt package available to you no matter which folder you are in, you may add the D:\MyMatlab\fcmclt to your MATLAB path. You may consult MATLAB documentation if you do not know how to do it.

Compatibility

The MEX files provided with the release have been generated using MATLAB R2022b on Windows x64 OS. The C code did not undergo any changes since MATLAB R13. The generation of MEX binaries has been tested most recently with 2018b, 2019b, and 2022b MATLAB releases.

Usage

The best way to learn how to use the functions provided in this package is to read their documentation. You may run >> help <function-name> to get the function description. In addition to the help text for the three main algorithms (fcmc, gkfcmc, and extfpm), you should consult the MATLAB source code for the test routines provided in .m files. That will provide you with examples of how to integrate these routines into your programs and how to initialize fuzzy partition matrix or evaluate the results.

The MATLAB source code may also provide references to additional reading material. The notation used in the source code is based primarily on this book:

• J.C. Bezdek, "Pattern Recognition with Fuzzy Objective Function Algorithms," Plenum Press, New York, 1981.

Support

The support is not provided for this package.

Implementation

The initial implementation of this package was done between 1992-1995. There were only a handful of minor changes done since that time. Those were limited to accomodate the MEX API changes introduced by MathWorks, Inc. over a long period of time (25+ years).

The code in ./src folder contains C and MATLAB sources. The C files are:

• extfpm.c: Extrapolates the fuzzy partition matrix using the presumed cluster centers
• fcmc.c: Implements the Fuzzy c-Means Clustering (FCMC) algorithm
• gkfcmc.c: Implements Gustafson-Kessel (GK) variant of the FCMC algorithm

The relevant MATLAB files are:

• Contents.m: Contains help text for the fcmclt package.
• fcmcinit.m: Very important routine that is used to generate the initial fuzzy partition matrix U₀.
• cltvalid.m: Calculates clustering validity functionals
• fscat.m: Calculates fuzzy scatter and covariance matrices for a fuzzy cluster
• testextfpm.m: Example of how to extrapolate fuzzy partition matrix from the presumed cluster centers
• testfcmc.m: Example of how to use FCMC algorithm
• testgk.m: Example of GK variant for two Gaussian classes
• testgk2.m: Example of GK for Gustafson's cross
• tstvalid.m: Example of how to use validity functionals for the FCMC algorithm

C code organization

The C functions include only two header files. The math.h and mex.h. The only two core MATLAB algorithms that are invoked within the C code are the matrix inverse and determinant.

For brevity, only the fcmc.c code structure is outlined here:

• The main function called by MATLAB is mexFunction()
  • Checks the input arguments and applies the defaults for optional ones as needed
  • Fetches pointers to vector data
  • Allocates some memory
  • Prepares for the FCMC routine
    • Selects the distance metric (may allocate additional memory and may invoke MATLAB inverse function in case of Mahalanobis metric)
  • Executes the FCMC algorithm (do_fcm())
  • Checks and creates output variables
• All other functions are written in plain C code

In essense, the mexFunction() is a MATLAB wrapper for the algorithms that are to be accelerated.

NOTE: If you would like to use the C code outside of MATLAB, you may need to do a few things:

• Make it re-entrant. That is, you would need to instantiate the variables that are currently global or static. That was not a problem for the MATLAB implementation, but may create problems if you try to integrate this code somewhere else.
• You would need to provide compatible implementations for matrix inverse and determinant MATLAB functions in case you have to use the parts of code that depend on these.
• Finally, you would need to replace the mexFunction() implementation with your own wrapper that would handle memory management, optional input arguments, etc. The memory management you will need to implement will likely differ from the way MATLAB manages memory. Especially when it comes to allocating and deallocating memory.

The code in gkfcmc.c and extfpm.c files is structured in a similar way.

Roadmap

There are no plans to expand this package with additional features.

Contributing

Pull requests will only be considered for the following contributions:

• Bug fixes (if you can find any)
• Interesting examples that show how to use the provided functions
  • Please try to provide a single self-contained MATLAB file that does not require any data files. If additional MATLAB toolboxes are required, please list them all.
  • Document your code! If your code cannot be easily reviewed it will be swiftly rejected no matter how great you think it is.

Authors

The software provided here has been developed by Bogdan Kosanovic in the early 1990s during his Ph.D. work at the University of Pittsburgh.

Acknowledgments

The author would like to acknowledge Dr. Brahim Hamadicharef who was at the time (August, 2003) with the University of Plymouth for submitting a fix to make the fcmclt package work for MATLAB 6.5 R13. If we are to trust the LinkedIn, Dr. Hamadicharef is now (2022) with the Institute of High Performance Computing (IHPC) in Singapore.

References

For general information on fuzzy clustering, partition coefficient (F), and partition entropy (H), the best starting point is

J.C. Bezdek, "Pattern Recognition with Fuzzy Objective
Function Algorithms", Plenum Press, New York, 1981.

For more information on nonfuzzy index (NFI), please refer to

M. Roubens, "Pattern Classification Problems and Fuzzy Sets",
Fuzzy Sets and Systems, 1:239-253, 1978.

For more information on minimum and mean hard tendencies (MinHT and MeanHT), please refer to

F.F. Rivera, E.L. Zapata, and J.M. Carazo, "Cluster validity
based on the hard tendency of the fuzzy classification",
Pattern Recognition Letters, 11:7-12, 1990.

For more information on minimum and maximum relative fuzziness (MinRF and MaxRF), please refer to

H.L. Gordon and R.L. Somorjai, "Fuzzy Cluster Analysis of
Molecular Dynamics Trajectories", Proteins: Structure,
Function, and Genetics, 14:249-264, 1992.

For more information on minimum nearest maximum membership cardinality (MinNMMcard) and related functionals including a way to go about the selection of fuzzy exponent m, please refer to

B.R. Kosanovic, "Signal and System Analysis in Fuzzy Information Space",
Ph.D. Dissertation, University of Pittsburgh, 1995.

For more information on compactness and separation index (S), please refer to

X.L. Xie and G. Beni, "A Validity Measure for Fuzzy Clustering",
IEEE Trans. PAMI, 13(8):841-847, 1991.

For more information on fuzzy hypervolume (Fhv), average partition density (Dpa), and partition density (Pd), please refer to

I. Gath and A.B. Geva, "Fuzzy clustering for the estimation of
the parameters of the components of mixtures of normal
distributions", Pattern Recognition Letters, 9:77-86, 1989.

License

MIT License