Fundamentals of Probabilistic Data Mining: 1.1 Lab Work

Master M2 MSIAM and MoSIG
Fundamentals of Probabilistic Data Mining

Graded lab and homeworks
http://chamilo.grenoble-inp.fr/courses/ENSIMAGWMM9AMO17/
1 Mixture models
The Unistroke alphabet, closely related to Graffiti1 , is an essentially single-stroke shorthand hand-
writing recognition system used in PDAs. The data set is composed of 50 × 6 time-trajectories
representing the drawing of letters A, E, H, L, O and Q in a plane.
Here you will focus on modelling letter A (actually drawn as a Λ). After some pre-processing,
we obtain the data set "Amerge.txt" (you can find it in the zip file), which is composed of every
stroke of every trial for the gestures associated with that letter (the temporal aspect of sequences
and the separations between sequences were lost here).
1.1 Lab work

1.1.1 Preparatory work and modelling
Do this before the class. Questions about this part will be answered only at the beginning of the
practical session.
1. Prove the reestimation formula for Gaussian Mixture Model (GMM) (exercise 2 in the slides).
2. Simulate a sample of size 500 of the following bivariate GMM:
0.3N (µ1 ; Σ1 ) + 0.7N (µ2 ; Σ2 )
with
−3 3 5 −2 5 2
µ1 = , µ2 = and Σ1 = Σ2 =
0 0 −2 1 2 2
Hint: numpy.random.multivariate_normal.
Plot the synthetic data set and check if it corresponds to the figures in the slides of the class
(Page 6).
3. Download (from chamilo), load and plot the Unistroke data set (letter A) and provide the
figure.
4. Do you think a 2-components GMM could be appropriate for letter A? Why?

1
http://en.wikipedia.org/wiki/Graffiti_(Palm_OS)
1
1.1.2 Data analysis: Gaussian model
1. Estimate a bivariate GMM on the letter A data set and provide the estimated parameters.
2. Label the data using the estimated model and show the pdf of the estimated GMM
(Provide one figure with the data labeled in color overlapping on the contours of the log(pdf),
please add inline labels for the contours)
Hint: mixture.GaussianMixture.predict, numpy.meshgrid.
3. To validate the assumption of bivariate Gaussian mixture:
(a) Plot each marginal histogram (in x and y) and add the estimated mixture of univariate
Gaussian pdfs to the figure.
(b) For each marginal, provide separate histograms of each cluster and add the estimated
univariate Gaussian pdf to the figure.
Hint: scipy.stats.norm.
4. Comment the results of questions 3 (a) and (b). What to think about the bivariate Gaussian
mixture assumption? Why?
5. Plot each data point xi with some colourmap corresponding to P (Zi = 1|Xi ) (you may plot
log P (Zi = 1|Xi ) instead). How to interpret that plot?
1.2 Mandatory additional questions
The aim of this part is to compare mixture of von Mises distributions with Gaussian mixtures.
1. Transform the Unistroke data to angular data. Plot the histogram of angles and comment.
2. Define von Mises and mixtures of von Mises distributions.
3. A priori, would a mixture of von Mises distributions be more or less adequate than Gaussian
mixtures on the real data set of part 1.1? Why?
4. Provide equations for the E-step and M-step of the EM algorithm for mixtures of von Mises
distributions. Justify these results with formal computations.
5. Fit a 2-components mixture of von Mises distributions on the Unistroke data set of part 1.1.
List the estimated parameters and color the data (in original form) by the estimated labels.
Hint: You may find an existing python library for mixtures of von Mises.
2
1.3 Optional additional questions
1. Consistent estimators of the number of components.
(a) Give a formal definition of consistent estimators of the number of components in a mix-
ture model. Write some state-of-the-art on that topic, choose one of the references
therein, justifying your choice. Provide a one-page description of the approach devel-
oped in that reference.
(b) Imagine, describe and implement a protocol to evaluate the consistency of any arbitrary
estimator of the number of components. Test this protocol on Gaussian mixtures to
check the consistency of that estimator.
2. Implementation of the mixtures of von Mises distributions
(a) Write your own sampling function and pdf function of Mixtures of von Mises distribu-
tions.
(b) Use your functions to simulate a 3-components mixture, with sample size of 1,000. Pro-
vide the figure showing the data colored by the true labels and the contour plot of the
log(pdf) of the simulated model (you may visualize them on 2D euclidean space).
(c) Estimate the parameters on the simulated data using your implementation. Comment
the results using parameters, histograms and bivariate plots with clusters (the same plot
as for (b) but using the estimated parameters).

Fundamentals of Probabilistic Data Mining: 1.1 Lab Work

Uploaded by

Copyright:

Available Formats

Fundamentals of Probabilistic Data Mining: 1.1 Lab Work

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fundamentals of Probabilistic Data Mining: 1.1 Lab Work

Uploaded by

Copyright:

Available Formats

Master M2 MSIAM and MoSIG

Fundamentals of Probabilistic Data Mining

1.1 Lab work

2. Simulate a sample of size 500 of the following bivariate GMM:

0.3N (µ1 ; Σ1 ) + 0.7N (µ2 ; Σ2 )

4. Do you think a 2-components GMM could be appropriate for letter A? Why?

3. To validate the assumption of bivariate Gaussian mixture:

1.2 Mandatory additional questions

2. Define von Mises and mixtures of von Mises distributions.

1. Consistent estimators of the number of components.

2. Implementation of the mixtures of von Mises distributions

You might also like