Adding a Classification attribute

[Tamir Basin]

Hi,

As shown in lesson 1.6, one can add a classification attribute.

I don't understand its benefits. When would I want to do that?

Cheers,

Tamir

[kirit ved]

hi,

class gives actual values while classification attribute gives the value decided by classifier. if both matches then it is correctly classified or mis classified.

[Tamir Basin]

Hi,

But the "class"is an ' attribute' already given as part of the data set.

Again, when or why would I add a 'classification' attribute (and how is it possible at all since we have to run the classifier first to come up with the results...)

Tamir

בתאריך יום שבת, 18 באפריל 2015 בשעה 06:47:16 UTC+3, מאת kirit ved:

[Ian Witten]

Well, you might never want to do that :-) — and you probably would never use it in the ordinary run of things.

But sometimes people do, and Weka provides a lot of facilities for experimentation, like this one.

For example, adding the classification would allow you to visualise the results of the machine learner — though of course you can already do that by right-clicking the results list, as you learned in Lesson 1.6. However, some classifiers produce probabilities, and you can arrange to add these probabilities, one for each class, instead of a single “classification” attribute. Then you could visualise the probabilities — something that you can’t do by right-clicking the results list.

Or how about this for an idea. You could run several different classifiers, add their classifications to the data set, and then use another classifier on top, that had access to the original attributes *and* the results of the other classifiers. That might be a way of getting good performance, and being able to add the classifications using different learners would allow you to investigate this. (Actually, this idea is called “stacking”, and Weka already has classifiers that do that, in the “meta” category: they’re called Stacking and StackingC. However, you might be able to come up with some other ways of using the classification that we haven’t thought of.)

But in general, the facility to add a classification attribute is not often used, and it’s OK to ignore it.

(and how is it possible at all since we have to run the classifier first to come up with the results...)

It’s done with a filter, a supervised attribute filter. The filter is called addClassification, and to use it you have to specify a particular classifier. If you like, you could try it! (Note: you have to set outputClassification to true in order for it to actually add the classification — or set outputDistribution if you want to add a probability distribution.)

hope this helps. Good question.

cheers

ian

Cheers,

Tamir

--
You received this message because you are subscribed to the Google Groups "WekaMOOC-general" group.
To unsubscribe from this group and stop receiving emails from it, send an email to wekamooc-gener...@googlegroups.com.
To post to this group, send email to wekamooc...@googlegroups.com.
Visit this group at http://groups.google.com/group/wekamooc-general.
To view this discussion on the web, visit https://groups.google.com/d/msgid/wekamooc-general/470304cc-bd75-48ab-a375-e0dcd8cc489d%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Tamir Basin]

Thank you Prof. :-)

Cheers,

Tamir

בתאריך יום ראשון, 19 באפריל 2015 בשעה 03:53:48 UTC+3, מאת Ian Witten:

[kirit ved]

hi,

Purpose of data mining is to use past data and predict for future. We may use past data where actual class is known. This is known as train data. Using this data classifier will build a model and apply this model on test data for which actual class is unknown or we want to predict that this can be done by using classification variable.

 prof. has already explained