New attribute created for PredictionModel subclass doesn't show up

colo

Hi,

I am currently stuck with a quite simple problem. I have created a subclass of PredictionModel and modified the method createPredictionAttributes() as follows:

protected Attribute createPredictionAttributes(ExampleSet exampleSet, Attribute label) {
	// create and add prediction attribute

	/* make the label attribute independent from an existing label in the example set (since this had to be generated as dummy when classifying documents)
	 * original code:
	 * Attribute predictedLabel = AttributeFactory.createAttribute(label, Attributes.PREDICTION_NAME);
	 * predictedLabel.clearTransformations();
	*/

	Attribute predictedLabel = AttributeFactory.createAttribute(Attributes.PREDICTION_NAME, Ontology.NOMINAL);		

	ExampleTable table = exampleSet.getExampleTable();
	Attributes attributes = exampleSet.getAttributes();

	table.addAttribute(predictedLabel);
	attributes.setPredictedLabel(predictedLabel);

	// check whether confidence labels should be constructed
	if (supportsConfidences(predictedLabel)) {
		Iterator<String> labelIterator = model.classIndex.iterator();
		while (labelIterator.hasNext()) {
			String labelValue = labelIterator.next();
			Attribute confidence = AttributeFactory.createAttribute(Attributes.CONFIDENCE_NAME + "(" + labelValue + ")", Ontology.REAL);
			table.addAttribute(confidence);
			attributes.setSpecialAttribute(confidence, Attributes.CONFIDENCE_NAME + "_" + labelValue);
		}

		// create an additional attribute for the confidence of the predicted value
		Attribute predictionConfidence = AttributeFactory.createAttribute(Attributes.CONFIDENCE_NAME, Ontology.REAL);
		table.addAttribute(predictionConfidence);
		attributes.setSpecialAttribute(predictionConfidence, Attributes.CONFIDENCE_NAME);
	}

	return predictedLabel;
}

Creating the "confidence(...)" attributes works fine. After this I want to have an additional attribute "confidence" that holds the confidence of the value that was actually predicted. When printing the attribute to console I get the index of the attribute, which is the number of the existing attributes + 1. But the additional attribute doesn't show up in my example set after "Apply Model". I tried using different names or adding it as regular attribute. No chance...

I would be glad about any help with this.

Best regards
Matthias

Marco_Boeck

Hi,

I'm sorry it took so long to answer, but it has been rather busy, and being ill didn't help anything either :-[
If you want to add attributes to an ExampleSet, you need to call these two methods:

exampleSet.getExampleTable().addAttribute(newAttribute);
exampleSet.getAttributes().addRegular(newAttribute);

Regards,
Marco

colo

Hi Marco,

I will let you get away with this answer because you were ill Did you even have a look at my code?
I have created new attributes pretty often, but in this case it simply isn't possible. I tried regular attributes as well as the shown attempt to add a special attribute. The strange thing is, that the attribute seems to be added (as printing out the attribute id indicates) but won't become visible. I guess there is some special context for the PredictionModel which is using a header example set, but I can't understand that adding lots of attributes for the confidences is working and an additional single one simply won't work...

Regards
Matthias

Marco_Boeck

Hi,

well yes, I did look at your code, but I may have misinterpreted the problem a bit :P
This is something I wrote a couple of weeks ago which is working just fine:

exampleSet.getExampleTable().addAttribute(newAttribute);
exampleSet.getAttributes().addRegular(newAttribute);
exampleSet.getAttributes().setSpecialAttribute(newAttribute, "YourSpecialAttributeName");

Regards,
Marco

colo

Hi Marco,

this works fine for me also - at any place and for any operator I modified or created except the PredictionModel subclass. I really can't understand this, since the "confidence(<labelValue>)" attributes are created properly.

I added this for testing

Attribute newAttribute = AttributeFactory.createAttribute("TestAttribut", Ontology.NOMINAL);
exampleSet.getExampleTable().addAttribute(newAttribute);
exampleSet.getAttributes().addRegular(newAttribute);
exampleSet.getAttributes().setSpecialAttribute(newAttribute, "YourSpecialAttributeName");
System.out.println(newAttribute);

and get

#16: TestAttribut (nominal/single_value)/values=[]

printed to the console. Before adding the attribute via exampleSet.getAttributes().addRegular(newAttribute); the id is -1, so adding it seems to work but it is not displayed!

I also tried to create this attribute if it's not existing from the other method performPrediction() implemented for this class, but it's the same behavior here.

It seems like I will have to add another workaround and create the attribute before the "Apply Model" operator is used. But I think you should test this in the special context given if you find the time for this.

Edit:
I just discovered another "funny" thing, when applying my workaround. I added "Generate Empty Attribute" and "Set Role" before using the "Apply Model" operator. Setting the attribute's values is done inside the performPrediction() method of the implemented model. But I only got missing values for the attribute. This was due to the code I initially posted, that should create this attribute. After commenting this out the correct values were assigned. So I must correct my opinion. Attempting to add the attribute is not only ignored it's also destroying existing attributes having the same name.  :-\


Regards
Matthias