rotemmairon/mvmdist

A Matlab package for probabilistic modeling of circular data with mixtures of von Mises distributions.

mvmdist

A Matlab package for probabilistic modeling of circular data with mixtures of von Mises distributions.

This package provides a class-based interface, similar to MATLAB's build-in functions for handling Gaussian mixture models. For further information on circular probability distributions and von Mises mixture models in particular, these papers give a comprehensive overview on the general concepts:

• J. Bentley (2006), "Modelling Circular Data using a Mixture of von Mises and Uniform Distributions"
• I. S. Dhillon and S. Sra (2003), "Modeling Data using Directional Distributions"

Installation

1. Clone or download this repository and add it to your MATLAB search path.
2. Compile the *.mex-file located at @VonMisesMixture/private by browsing to that directory in MATLAB and running mex sampleVonMisesMex.c from the command line.

Examples

Constructing a von Mises mixture model

The following code creates a von Mises mixture model with two mixture components.

p = [0.5; 0.5];         % Mixture weights.
mu = [-pi/2; pi/2];     % Component means.
kappa = [5; 10];        % Concentration parameters of components.

vmm = VonMisesMixture(p, mu, kappa);

Plotting the probability density function

Given a vector of angular values, the probability density function of a von Mises mixture model can be computed with the pdf() method provided by the VonMisesMixture class. The following code example plots the probability density function of a VonMisesMixture class instance.

angles = linspace(-pi, pi, 1000)'; % The pdf() function expects a column-vector as input.
likelihoods = vmm.pdf(angles);
plot(angles, likelihoods); grid on;

Drawing samples from a von Mises mixture model

This implementation uses the method introduced by Barabesi (2005) to generate samples from a von Mises distribution. To speed up the sampling process, a mex-function is used by default which is significantly faster than the plain MATLAB implementation (especially for a large number of samples).

nSamples = 10000;
samples = vmm.random(nSamples); % Generate samples using *.mex-function.

If the plain MATLAB sampling method should be used, a second argument has to be passed to the random() function which sets the internal useMex flag to false:

samples = vmm.random(nSamples, false); % Generate samples using MATLAB implementation.

Maximum likelihood parameter estimation

Parameter estimation is conducted via an Expectation-Maximization (EM) approach described in Hung et al. (2012). The following example shows how a model can be fitted on a set of data samples drawn from an initial von Mises mixture distribution with three components.

p = [0.3; 0.4; 0.3];         % Mixture weights.
mu = [-pi/2; 0; pi/3];       % Component means.
kappa = [5; 10; 2.5];        % Concentration parameters of components.

vmm = VonMisesMixture(p, mu, kappa);  % Initialize model.
samples = vmm.random(10000);          % Draw 10000 samples.

% Fit new model on data samples assuming 3 mixture components.
nComponents = 3;
fittedVmm = fitmvmdist(samples, nComponents, ...
  'MaxIter', 250); % Set maximum number of EM iterations to 250

% Plot initial and fitted distributions.
angles = linspace(-pi, pi, 1000)';
likelihoodsInitial = vmm.pdf(angles);
likelihoodsFitted = fittedVmm.pdf(angles);

plot(angles, likelihoodsInitial); hold on;
plot(angles, likelihoodsFitted); grid on;
axis([-pi, pi, 0, 1]);