How to set parameter of SMO?

IngoRM

Original message from SourceForge forum at http://sourceforge.net/forum/forum.php?thread_id=2038567&;forum_id=390413

I want to use SMO to get well-calibrated probability
I call the "setparameter" method in my program,but the output's probability is still equal to 0 or 1, i.e. it doesn't work.(I call the method getParameter("M"), the result is "true"! , not default value "false")
But I run SMO by GUI, I can set parameter to get well-calibrated probability

I also try to set others parameters of SMO or other Weka algorithm. it is also invalid. But I can set parameter in non-weka algorithm, it's valid

Someone please help me...

My Java Code: 
//build model
Operator SMOOperator = OperatorService.createOperator("W-SMO"); 
SMOOperator.setParameter("M", "true");
Learner learner = (Learner) SMOOperator;
Model model = learner.learn(trainingset);

// apply model
IOContainer testingSetContainer = testingSetContainer.append(model);
Operator modelApp = OperatorService.createOperator(ModelApplier.class);
testingSetContainer = modelApp.apply(testingSetContainer);

// print results
ExampleSet resultSet = resultSetContainer.get(ExampleSet.class); 
ExampleTable tableReader = resultSet.getExampleTable();
DataRowReader rowReader = tableReader.getDataRowReader();
DataRow row; 

Attribute aid = resultSet.getAttributes().getId();
Attribute apl = resultSet.getAttributes().getPredictedLabel();
Attribute al = resultSet.getAttributes().getLabel();
Attribute ac = resultSet.getAttributes().get("confidence(1)");

while(rowReader.hasNext())
{
row=(DataRow)rowReader.next();
System.out.println("id = "+row.get(aid)+" label = "+row.get(al)+" prediction = "+row.get(apl)+" confidence = "+row.get(ac));
}

Result:
id = 685.0 label = 0.0 prediction = 0.0 confidence = 0.0
id = 701.0 label = 0.0 prediction = 0.0 confidence = 0.0
id = 697.0 label = 0.0 prediction = 0.0 confidence = 0.0
.........................................................


Answer by Harri:

In SMO, you have to set the logistic parameter (I forget exact name but there is only one) 'true' to get true probabilities. 

Harri


Answer by topic starter:

I have been checked the tutorial. 
In SMO, the description of parameter "M" is 

M: Fit logistic models to SVM outputs. (boolean; default: false)

I am sure that I set the "correct" parameter, correct parameter name and correct value
(if I type wrong name ex. "m"(lowercase) or wrong value ex. "trrue" then the error information will print out)

Why I set "M" parameter in RapidMiner GUI is work, but in java code is wrong?

I also try to set other parameter value randomly, such as 
C: The complexity constant C. (default 1) (real; -1-+1) 
L: The tolerance parameter. (default 1.0e-3) (real; -1-+1)

My accuracy is not change. it seems that I can only use default setting
This problem also exist in Weka learning algorithm "W-MultilayerPerceptron", 
but not exist in RapidMiner learning algorithm "DecisionTree" 

this problem is very confused.


Answer by Ingo Mierswa:

Hello,

does it also apply if you use

Model model = SMOOperator.apply(new IOContainer(trainingset)).getInput(Model.class);

instead of 

Learner learner = (Learner) SMOOperator; 
Model model = learner.learn(trainingset); 

Probably the parameters are not given to the Weka operators in the learn(...) method.

Cheers,
Ingo


Answer by topic starter:

Hello,

the line of code will occur compile error...

Model model = SMOOperator.apply(new IOContainer(trainingset)).getInput(Model.class);

In IOContainer class, getInput() method is not exist
I try to revise these code like this...

Operator SMOOperator = OperatorService.createOperator("W-SMO"); 
SMOOperator.setParameter("M", "true"); 
try {
Model model = SMOOperator.apply(trainingSetIOContainer).get(Model.class);
} catch (OperatorException e1) {
e1.printStackTrace();
}

But the parameter value which I set is also invalid, the problem still exist in Weka learning algorithm

Do someone know other way to set parameter value for weka operator?


Answer by Ingo Mierswa:

Ok, I found the reason: since all Weka operators are dynamically created the parameters are not initialized directly after operator creation and you have to do this manually. Just use the following lines

Operator smoOperator = OperatorService.createOperator("W-SMO");
smoOperator.getParameterTypes(); // <-- this is the important workaround
smoOperator.setParameter("M", "true");

for operator creation and parameter setting and it will work. I will also add a fix to the CVS version of RapidMiner so that you will not have to invoke "getParameterTypes()" for the Weka operators for future versions of RapidMiner.

Hope that helps. Cheers,
Ingo