You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mahout.apache.org by is...@apache.org on 2013/11/21 12:07:08 UTC
svn commit: r1544109 -
/mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext
Author: isabel
Date: Thu Nov 21 11:07:08 2013
New Revision: 1544109
URL: http://svn.apache.org/r1544109
Log:
MAHOUT-1245 - fixing images
Modified:
mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext
Modified: mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext
URL: http://svn.apache.org/viewvc/mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext?rev=1544109&r1=1544108&r2=1544109&view=diff
==============================================================================
--- mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext (original)
+++ mahout/site/mahout_cms/trunk/content/users/clustering/dirichlet-process-clustering.mdtext Thu Nov 21 11:07:08 2013
@@ -1,6 +1,7 @@
Title: Dirichlet Process Clustering
+
<a name="DirichletProcessClustering-Overview"></a>
-# Overview
+# Overview Dirichlet process clustering
The Dirichlet Process Clustering algorithm performs Bayesian mixture
modeling.
@@ -99,10 +100,8 @@ Invocation using the command line takes
-a0 <the alpha_0 parameter to the Dirichlet Distribution>
-x <maximum number of iterations> \
-k <number of models to create from prior> \
- -md <the ModelDistribution class name. Default NormalModelDistribution>
-\
- -mp <the ModelPrototype class name. Default
-SequentialAccessSparseVector> \
+ -md <the ModelDistribution class name. Default NormalModelDistribution> \
+ -mp <the ModelPrototype class name. Default SequentialAccessSparseVector> \
-dm <optional DistanceMeasure class name for some ModelDistribution>
-ow <overwrite output directory if present>
-cl <run input vector clustering after computing Clusters>
@@ -171,9 +170,9 @@ The points are generated as follows:
In the first image, the points are plotted and the 3-sigma boundaries of
their generator are superimposed. It is, of course, impossible to tell
which model actually generated each point as there is some probability -
-perhaps small - that any of the models could have generated every point.
+perhaps small - that any of the models could have generated every point
-!SampleData.png!
+
In the next image, the Dirichlet Process Clusterer is run against the sample points using a NormalModelDistribution with m=\[0.0, 0.0\](0.0,-0.0\.html)
sd=1.0. This distribution represents the least amount of prior
@@ -190,7 +189,7 @@ As Dirichlet clustering is an iterative
. These illustrate the cluster convergence process over the last several
iterations and can be helpful in tuning the algorithm.
-!DirichletN.png!
+
The next image improves upon this situation by using a
SampledNormalDistribution. In this distribution, the prior models have
@@ -201,14 +200,14 @@ different pdf for each point and the ite
more-likely models given this value. The result is a decent capture of the
sample data parameters but there is still some over-fitting.
-!DirichletSN.png!
+
The above image was run through 20 iterations and the cluster assignments
are clearly moving indicating the clustering is not yet converged. The next
image runs the same model for 40 iterations, producing an accurate model of
the input data.
-!DirichletSN40.png!
+
The next image uses an AsymmetricSampledNormalDistribution in which the
model's standard deviation is also represented as a 2-d vector. This causes
@@ -218,7 +217,7 @@ the actual sample data quite well. Had w
generated in a similar manner then this distribution would have been the
most logical model.
-!DirichletASN.png!
+
In order to explore an asymmetrical sample data distribution, the following
image shows a number of points generated according to the following
@@ -231,14 +230,14 @@ parameters. Again, the generator's 3-sig
* 300 samples m=\[0.0, 2.0\](0.0,-2.0\.html)
sd=\[0.1, 0.5\]
-!AsymmetricSampleData.png!
+
The following image shows the results of applying the symmetrical
SampledNormalDistribution to the asymmetrically-generated sample data. It
does a valiant effort but does not capture a very good set of models
because the circular model assumption does not fit the data.
-!2dDirichletSN.png!
+
Finally, the AsymmetricSampledNormalDistribution is run against the
asymmetrical sample data. Though there is some over-fitting, it does a
@@ -248,8 +247,8 @@ slightly different results. Compare the
for 20 iterations with another run of numClusters=40 models for 40
iterations.
-!2dDirichletASN.png!
-!2dDirichletASN4040.png!
+
+
<a name="DirichletProcessClustering-References"></a>
# References
@@ -263,4 +262,4 @@ model is found.
The Neal and Blei references from the McCullagh and Yang paper are also
good. Zoubin Gharamani has some very [nice tutorials out which describe why non-parametric Bayesian approaches to problems are very cool](http://learning.eng.cam.ac.uk/zoubin/talks/uai05tutorial-b.pdf)
-, there are video versions about as well.
+, there are video versions about as well.
\ No newline at end of file