[SOLVED] Weka "select attributes" from Linux terminal

Issue

I am interested in including some results that come from Weka in a system. I tried to look for similar functions in Rweka and in python, but I could not so I think that running the process from a Linux terminal would solve my problem. In the "select attributes" pannel, I chose:

Attribute evaluator: CfsSubsetEval -P 1 -E 1 Search Method: MultiObjectiveEvolutionarySearch -generations 20 -population-size 100 -seed 1 -algorithm 0 -report-frequency 20 -log-file /home/aurorax/Git_repos/postdoc/QoI/to_submit/weka-3-8-5

When I run it, it shows the following "code":

=== Run information ===

Evaluator:    weka.attributeSelection.CfsSubsetEval -P 1 -E 1
Search:       weka.attributeSelection.MultiObjectiveEvolutionarySearch -generations 20 -population-size 100 -seed 1 -algorithm 0 -report-frequency 20 -log-file /home/aurorax/Git_repos/postdoc/QoI/to_submit/weka-3-8-5
Relation:     metricsAndRMSE1-weka.filters.unsupervised.instance.RemoveDuplicates
Instances:    293
Attributes:   12
              qcomp
              qpla
              qout
              qoutMean
              qoutMedian
              qprob
              qrec
              qinvMean
              qinvMedian
              qBMEmean
              qBMEmedian
              rmse1
Evaluation mode:    evaluate on all training dat

And at the end of it, it shows what I am interested in, that is on what attributes should be selected:

Selected attributes: 1,3,11 : 3
                     qcomp
                     qout
                     qBMEmedian

Could you please provide me some guidance? I already checked some of the documentation, but could not find a way to do this.

Thank you very much!

Solution

Yes, you can run evaluators from the command-line, though not commonly used.

You can query the evaluator's help like this, e.g., for CfsSubsetEval:

java -cp weka.jar weka.attributeSelection.CfsSubsetEval -h

The output is split into General options and ones specific to CfsSubsetEval. The general ones are:

General options:

-h
    display this help
-i <name of input file>
    Sets training file.
-c <class index>
    Sets the class index for supervised attribute
    selection. Default=last column.
-s <class name>
    Sets search method for subset evaluators.
-x <number of folds>
    Perform a cross validation.
-n <random number seed>
    Use in conjunction with -x.

Assuming that your dataset is located in /your/dataset.arff and the class attribute is the last one, you get the following:

java -cp weka.jar weka.attributeSelection.CfsSubsetEval \
  -i /your/dataset.arff \
  -c last \
  -s "SEARCH_METHOD_CLASSNAME + OPTIONS" \
  [other CfsSubsetEval options]

In order to ensure that all packages get loaded correctly, you should use the weka.Run class to launch your attribute selection. This gives us:

java -cp weka.jar weka.Run \
  weka.attributeSelection.CfsSubsetEval \
  -i /your/dataset.arff \
  -c last \
  -s "SEARCH_METHOD_CLASSNAME + OPTIONS" \
  [other CfsSubsetEval options]

Or, in your case:

java -cp weka.jar weka.Run \
  weka.attributeSelection.CfsSubsetEval \
  -i /your/dataset.arff \
  -c last \
  -s "weka.attributeSelection.MultiObjectiveEvolutionarySearch -generations 20 -population-size 100 -seed 1 -algorithm 0 -report-frequency 20 -log-file /home/aurorax/Git_repos/postdoc/QoI/to_submit/weka-3-8-5" \
  -P 1 -E 1

Final note: Be careful to escape double quotes and backslashes correctly. If your search method command-line should contain them (not in the case that you posted), you need to escape them with backslashes.

For example:

weka.attributeSelection.FunkySearch -b "some option \"with nested bits\""

Will change into this as part of the command-line:

...
-s "weka.attributeSelection.FunkySearch -b \"some option \\"with nested bits\\"\""
...

Answered By - fracpete