/pltm-east

Pouch latent tree models for multidimensional clustering

Primary LanguageJava

pltm-east

This directory contains the Java implementation of the PLTM algorithm described in the paper "Variable selection in model-based clustering: To do or to facilitate" (ICML 2010). It also contains the sub-directory data where data sets in ARFF format can be found.

The directory includes shell scripts for running on Unix machines.

To learn PLTM (named output.bif) on a data set (e.g. data/iris.arff):

$ ./run.sh data/iris.arff

To show the options of PLTM-EAST, type run.sh without any parameters:

$ ./run.sh 
usage: PltmEast [OPTION] data_file
 -c,--class <class_variable>       specify the zero-based index of class
                                   variable, or none, first, last
                                   (default: last)
 -i,--initial-model <model_file>   start the search from an initial model
 -m,--allow-missing                allow missing data
 -o,--output-file <output_file>    specify the output BIF file (default:
                                   output.bif)
 -s,--setting <setting_file>       use the specified settings file
                                   (default: settings.xml)

To specify the output file name:

$ ./run.sh -o iris.bif data/iris.arff

To evaluate a model using NMI:

$ ./evaluateNMI.sh data/iris.arff iris.bif

This script also accepts multiple model files, such as:

$ ./evaluateNMI.sh data/iris.arff iris*.bif

In case other measures are used for evaluating the clusterings, the marginal probabilities of the latent variables for every data cases can be computed by:

$ ./classify.sh data/iris.arff iris.bif

It will generate several CSV files, each of which contains the marginal probabilities of a latent variable for every data case. It will also generate a CSV file for the class variable.

Compile

To compile the source code and create the distribution directory, type the following command in the project directory:

$ ant dist

The distribution directory can be found in the dist directory.

To compile it using Eclipse, you will need to install the JavaCC Eclipse Plug-in.

Enquiry

For any inquiries, please email kmpoon@eduhk.hk.

References