Nabble - WEKA

dimensionality reduction using PrincipalComponents

2009-11-25T19:32:42Z

Hi,

I am new to Weka, and I am trying to learn how to use it. I see in the
documentation of PrincipalComponents.java:

Performs a principal components analysis and transformation of the data.
Use in conjunction with a Ranker search. Dimensionality reduction is
accomplished by choosing enough eigenvectors to account for some
percentage of the variance in the original data---default 0.95 (95%).
Attribute noise can be filtered by transforming to the PC space,
eliminating some of the worst eigenvectors, and then transforming back
to the original space.

Great! I want to try to do this. I wrote a program to attempt this on a small
geometric example (code attached). I create a training set that consists of
two (x, y) pairs representing the line y=2x. I build my principal components.
I then transform a point from a test set into the principal components space
and then back. If I use the point (2, 4), I would expect to get exactly the
same point back. If I were to use the point (2, 4.5), I would expect it to be
projected onto the line y=2x (i.e. into the principal components space), and
then transformed back into the original coordinate space yielding a point
somewhere on the line y=2x between (2, 4) and (2.25, 4.5). This is not what
is happening. (2, 4) is getting transformed to (3, 3). I also
noticed that the
eigenvectors for any 2x2 matrix are (+/-.7072, +/-.7072) regardless of the
elements in the matrix. Is that correct?

Does my code need to be modified, or is it my expectations that need to be
modified?

Thanks in advance!

David

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Main.java (2K) Download Attachment

Attribute Selection

2009-11-25T17:57:11Z

Hi All, I have a datset which is too big for to use directly in Weka and i would like to choose say a subset of 200 from my dataset. How can i use weka or excel to select the best 200 from my dataset of over a thousand. Thanks

Re: Principal Component Analysis

2009-11-25T17:11:35Z

> I have to two different data sets, one for training and one for test,
> concerning with a classification problem.
> I plan to use PCA for reducing feature dimensions.
> After applying PCA to the training set, I got a list of new features as
> follows:
>
> Ranked attributes:
> 0.9515    1 -0.327w20-0.327w18-0.327w21-0.324w19-0.256w4...
> 0.9037    2 -0.284w8-0.23w69-0.23w68-0.23w70+0.222w4...
> 0.858     3 -0.309w22-0.257w2+0.242w48+0.242w47+0.242w49...
> 0.8152    4 -0.237w54-0.223w57+0.216w59-0.212w48-0.212w47...
> ...
>
> I have questions:
> How to transform both training and test sets by using the new features ?

Use "batch-filtering" with the following filter:
weka.filters.supervised.attribute.AttributeSelection

And read FAQ "How do I generate compatible train and test sets that
get processed with a filter?". Link to the FAQs available from the
Weka homepage.

> Is it possible to do this task with GUI of weka 3.4.13 ?

No version of the Explorer allows this, not sure about the
Knowledgeflow. You normally do that from command-line (SimpleCLI or
terminal).

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Principal Component Analysis

2009-11-25T16:50:35Z

Dear All,

I have to two different data sets, one for training and one for test,
concerning with a classification problem.
I plan to use PCA for reducing feature dimensions.
After applying PCA to the training set, I got a list of new features as follows:

Ranked attributes:
0.9515    1 -0.327w20-0.327w18-0.327w21-0.324w19-0.256w4...
0.9037    2 -0.284w8-0.23w69-0.23w68-0.23w70+0.222w4...
0.858     3 -0.309w22-0.257w2+0.242w48+0.242w47+0.242w49...
0.8152    4 -0.237w54-0.223w57+0.216w59-0.212w48-0.212w47...
...

I have questions:
How to transform both training and test sets by using the new features ?
Is it possible to do this task with GUI of weka 3.4.13 ?

Best regards,
Peerasak

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: jython - prediction with deserialized model

2009-11-25T13:10:37Z

> i try using a serialized classifier with jython ,but getting no success.

> What could be my mistake?
>
> Many thanks ,
> Christian
>
> import sys
> import weka.core.SerializationHelper as pmodel
> import weka.classifiers.Classifier.classifyInstance as classify
> import java.io.FileReader as FileReader
> import weka.core.Instances as Instances
>
> model=pmodel.read("/weka-3-7-0/data/nb_iris.model")
> file = FileReader("/weka-3-7-0/data/iris.arff")
>
> data=Instances(file)
> data.setClassIndex(data.numAttributes()-1)
>
> print data.numAttributes()
> print data.instance(98)
>
> cl=classify(model,data.instance(98))
>
>
> Output:
> 5
> 5.1,2.5,3,1.1,Iris-versicolor
> Traceback (most recent call last):
> File "scoring.py", line 23, in <module>
> cl=classify(model,data.instance(98))
> at java.util.ArrayList.RangeCheck(ArrayList.java:547)
> at java.util.ArrayList.get(ArrayList.java:322)

Since I don't know what your "classify" method does (that seems to be
the source of the error), I've attached an example of mine that
outputs the predictions made with a serialized model.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

[model.py]

import sys import weka.core.SerializationHelper as pmodel import weka.classifiers.Classifier.classifyInstance as classify import java.io.FileReader as FileReader import weka.core.Instances as Instances import weka.core.converters.ConverterUtils.DataSource as DSource model = pmodel.read("/home/fracpete/temp/nb.model") file = "/home/fracpete/temp/iris.arff" data = DSource.read(file) data.setClassIndex(data.numAttributes() - 1) for i in xrange(0, data.numInstances()): inst = data.instance(i).copy() inst.setClassMissing() pred = model.classifyInstance(inst) label = data.classAttribute().value(int(pred)) print str(i) + ". " + str(pred) + "/" + label + ": " + str(data.instance(i))

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: RE: weka classifiers

2009-11-25T12:31:36Z

Please no top-posting, see mailing list etiquette why
(http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html). In
the future I'll refuse to answer questions that don't adhere to the
etiquette.

> Thank you Peter for your reply. Could you please tell me whether it does
> classification/regression when it receives different types of features or it
> converts the numeric ones into nominal and performs classification each
> time?

There's different classifiers for regression and classification. Some
of them can do both.

> Does weka always perform binary classification only?

Once again, it depends on the classifier.

Rule of thumb: if a classifier cannot handle the data, it will tell you.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

RE: weka classifiers

2009-11-25T12:28:24Z

Hi All,

Thank you Peter for your reply. Could you please tell me whether it does classification/regression when it receives different types of features or it converts the numeric ones into nominal and performs classification each time?

Does weka always perform binary classification only?

Many thanks,

Nancy

> From: fracpete@...
> Date: Wed, 25 Nov 2009 15:00:34 +1300
> Subject: Re: [Wekalist] weka classifiers
> To: wekalist@...
>
> > Can someone please tell me based on which features, weka classifiers perform
> > the classification? Do i determine specific feature/s first or I should
> > write something in the dataset or put the features in specific order?
>
> On the commandline, you define via the -c parameter which attribute is
> to be the class attribute. In the Explorer, you select the appropriate
> attribute from the combobox on the preprocess panel.
> The classifier will evaluate (what exactly happens here depends on the
> actual classifier) all other attributes against this class attribute
> then.
>
> Cheers, Peter
> --
> Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
> http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174
>
> _______________________________________________
> Wekalist mailing list
> Send posts to: Wekalist@...
> List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
> List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Windows Live: Make it easier for your friends to see what you’re up to on Facebook.
_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Extract Rules from trained Artificial Neural Network

2009-11-25T11:43:45Z

> Is there a way to extract rules from a trained a Artificial Neural Network
> using Weka?

Not that I know of.

> If there isn't, how can I develop this feature and share with the main weka
> package?

Read FAQ "How can I contribute to WEKA?". Link to the FAQs available
from the Weka homepage.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Iterating through instances

2009-11-25T11:07:26Z

Curtis Jensen wrote:

>
> I'm also confused what the int parameter is for. If each individual
> instance has a list of attributes, why does that attribute's value
> method take an integer index? Isn't which attribute already
> identified by the actual Attribute object?
>

I think you are getting confused between Instance.value(int) and
Attribute.value(int). The former, as Peter said, returns you the value
of an attribute in internal format. This is a primitive double. For
numeric/date attributes it *is* the value of the attribute. For nominal
attributes it is an *index* coded as a double. This is where
Attribute.value(int) comes in. If you want to know the string value that
corresponds to that index for nominal/string attributes then you need to
look up the value by calling Attribute.value(index_of_att_value), where
index_of_att_value is the double (cast to an int) you retrieved from
Instance.value(index_of_attribute).

Alternatively, as Peter said, you can just call
Instance.stringValue(index_of_attribute) for nominal/string attribute
*only*. This goes straight to the string representation and avoids
having to do the Attribute.value(index_of_att_value) lookup.

Cheers,
Mark.

--
Mark Hall
Senior Developer/Consultant, Pentaho Open Source Business Intelligence
Citadel International, Suite 340, 5950 Hazeltine National Dr., Orlando,
FL 32822, USA
+64 7 348-7099 office, +64 21 399-132 mobile, +1 815 550-8637 fax,
Skype: mark.andrew.hall, Yahoo: mark_andrew_hall
Download the latest release today
<http://www.sourceforge.net/projects/pentaho>

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: problem in using svm with weka

2009-11-25T10:23:19Z

> I have problem using SVM classifer and would be very grateful to you if you
> could help me.
>
> First when I choose libsvm function, The following message is displayed:
> « Problem evaluating classifier :libsvm classes not in classpath! ».
> what does that mean ?

Read FAQ "How do I use libsvm in WEKA?". Link to the FAQs is available
from the Weka homepage.

> Second,I succeded to run SMO function, but the displayed results are not
> clear for me and I can not undesrtand the results
> Is there any file/text/paper/url explaining the SMO output ?

Read Platt's paper about SMO. Paper reference should be in the Javadoc for SMO.

> Third , I would like to know how can I use the learning results for the
> testing phase (evaluation)
> How can I include the test file (Do I have to put the class of each line ?)

That should be rather obvious. In the Explorer, you just use "Supplied
test set" button. On the command-line, using the "-h" for displaying
the help will tell you that you can use the "-T" option to supply a
test set.

And yes, you have to include the class for each row in the data. How
else is Weka supposed to determine whether the prediction by the built
classifier was correct or not?

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Save Attribute Stats

2009-11-25T10:19:11Z

> In the "Preprocess" tab of the Weka Explorer, many stats on the
> attributes are displayed. Is there a way to save these stats (count
> and distribution numbers) to a text file?

Not that I know of. The weka.core.Instances class offers the method
"attributeStats(int)", which returns a weka.core.AttributeStats
object, containing all relevant information.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Iterating through instances

2009-11-25T10:14:57Z

[...]

> I'm working with nominal and numeric values only.
> My JavaDoc for the value method says:
> Returns a value of a nominal or string attribute. Returns an empty
> string if the attribute is neither a string nor a nominal attribute.

Not sure what version of Weka you're using. I'm using 3.7.0ish.

> >From that, it sounds like the value method will not work for numeric attributes?

Correct. The method "stringValue" is for attribute types that are
normally represented by a string, e.g., NOMINAL, STRING, DATE (also
RELATIONAL).

> I'm also confused what the int parameter is for. If each individual
> instance has a list of attributes, why does that attribute's value
> method take an integer index? Isn't which attribute already
> identified by the actual Attribute object?

The "int" parameter is the index of the attribute in the dataset,
which provides a fast and simple access (I prefer this method). Also,
if the Attribute object that you're using is not part of the Instances
object that you're manipulating, the access will fail (it's just a
simple object reference check that is used).

The ARFF format does not enforce unique attribute names. If you try to
return an Attribute object by name, you will always only get the first
one. Hence, the access via the "int" index is the safer one, IMHO.

[...]

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: How to evaluate a new boosting algorithm

2009-11-25T10:07:57Z

> I've finished the implementation of a new Boosting algorithm for classification problem using the regression REPTree as weak leaner. To evaluate the new classifier, I deleted all class values of the training data set to create the testing data set, then I used the class Evaluation to evaluate my classifier with the new testing data set. In comparison with the training data, I can see that all instances of the testing data set are well classified (100% of instances is well classified). Is this enough to confirm that my classifier is good ?

Evaluating on the training data isn't really a good test.

> Is there any other method to evaluate a new classifier ?

Set up an experiment in the Experimenter: 1. lots of datasets with
varying size, 2. compare it against other standard algorithms
(NaiveBayes, J48, SMO, AdaBoostM1, MultiBoostAB, Stacking, ...)

> --- On Wed, 11/25/09, Peter Reutemann <fracpete@...> wrote:
>
> From: Peter Reutemann <fracpete@...>
> Subject: Re: [Wekalist] Iterating through instances
> To: "Weka machine learning workbench list." <wekalist@...>
> Date: Wednesday, November 25, 2009, 3:35 PM
>
> > I'm attempting write a clusterer for Weka.

Do you think, anybody is interested in this unrelated information? New
question, new post (and no left-overs from other posts, of course).

[...]

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

problem in using svm with weka

2009-11-25T10:07:40Z

Dear Weka members,

I have problem using SVM classifer and would be very grateful to you if you could help me.

First when I choose libsvm function, The following message is displayed: « Problem evaluating classifier :libsvm classes not in classpath! ».

what does that mean ?

Second,I succeded to run SMO function, but the displayed results are not clear for me and I can not undesrtand the results

Is there any file/text/paper/url explaining the SMO output ?

Third , I would like to know how can I use the learning results for the testing phase (evaluation)

How can I include the test file (Do I have to put the class of each line ?)

Many thanks in advance for your help

**************************************************

Mohamed Mahdi Boudabbous

Master student

Faculty of Economic Sciences and management of Sfax
LARIS - MIRACL Laboratory

Address : FSEGS, BP 1088, 3018 Sfax, Tunisia

**************************************************

Save Attribute Stats

2009-11-25T09:49:36Z

In the "Preprocess" tab of the Weka Explorer, many stats on the
attributes are displayed. Is there a way to save these stats (count
and distribution numbers) to a text file?

Thanks,
Curtis

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Iterating through instances

2009-11-25T09:40:32Z

On Wed, Nov 25, 2009 at 12:35 AM, Peter Reutemann <fracpete@...> wrote:

>> I'm attempting write a clusterer for Weka.
>> I'm having trouble getting the individual attribute values for a data
>> instance. I'm handling numeric and nominal types.
>>
>> In going through the Weka code and online documentation, I'm confused
>> on how to get an individual value from the data instances.
>
> What kind of value do you want to have? Internal? String representation?
>
> "value(Attribute/int)" returns the actual value for numeric attributes
> and for nominal/string/relational the index. Copy/paste from the
> Javadoc:
> "the specified value as a double (If the corresponding attribute is
> nominal (or a string) then it returns the value's index as a double)."
>
> If you want to retrieve the actual "string" representation of
> non-numeric attributes (that includes Date attributes!), use
> "stringValue(Attribute/int)". Copy/paste from the Javadoc:
> "Returns the value of a nominal, string, date, or relational attribute
> for the instance as a string. "
>

I'm working with nominal and numeric values only.
My JavaDoc for the value method says:
Returns a value of a nominal or string attribute. Returns an empty
string if the attribute is neither a string nor a nominal attribute.

>From that, it sounds like the value method will not work for numeric attributes?

I'm also confused what the int parameter is for. If each individual
instance has a list of attributes, why does that attribute's value
method take an integer index? Isn't which attribute already
identified by the actual Attribute object?

> Most algorithms just use the index of nominal/string/relational
> attributes internal and not the string representation (to save space),
> i.e., using the internal format retrieved via "value(Attribute/int)".
>
>> See code
>> below:
>>
>>
>> int numAttrs = data.numAttributes();
>> int numInstances = data.numInstances();
>>
>> for (int instIdx = 0; instIdx < numInstances; instIdx++) {
>> Instance currInst = data.instance(instIdx);
>>
>> for (int attrIdx = 0; attrIdx < numAttrs; attrIdx++) {
>> Attribute currAttr = currInst.attribute(attrIdx);
>> if (currAttr.isNominal()) {
>> // now what?
>> }
>> else if (currAttr.isNumeric()) {
>> // now what?
>> }
>> }
>> }
>
>
> Cheers, Peter
> --
> Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
> http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174
>
> _______________________________________________
> Wekalist mailing list
> Send posts to: Wekalist@...
> List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
> List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html
>

jython - prediction with deserialized model

2009-11-25T09:39:48Z

Hi,

i try using a serialized classifier with jython ,but getting no
success. What could be my mistake?

Many thanks ,
Christian

import sys
import weka.core.SerializationHelper as pmodel
import weka.classifiers.Classifier.classifyInstance as classify
import java.io.FileReader as FileReader
import weka.core.Instances as Instances

model=pmodel.read("/weka-3-7-0/data/nb_iris.model")
file = FileReader("/weka-3-7-0/data/iris.arff")

data=Instances(file)
data.setClassIndex(data.numAttributes()-1)

print data.numAttributes()
print data.instance(98)

cl=classify(model,data.instance(98))

Output:
5
5.1,2.5,3,1.1,Iris-versicolor
Traceback (most recent call last):
File "scoring.py", line 23, in <module>
cl=classify(model,data.instance(98))
at java.util.ArrayList.RangeCheck(ArrayList.java:547)
at java.util.ArrayList.get(ArrayList.java:322)
at weka.core.Instances.attribute(Instances.java:326)
at
weka.classifiers.bayes.NaiveBayes.distributionForInstance(NaiveBayes.java:362)
at
weka.classifiers.AbstractClassifier.classifyInstance(AbstractClassifier.java:72)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at
sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:39)
at
sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:25)
at java.lang.reflect.Method.invoke(Method.java:597)

java.lang.IndexOutOfBoundsException:
java.lang.IndexOutOfBoundsException: Index: 0, Size: 0

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Extract Rules from trained Artificial Neural Network

2009-11-25T07:06:38Z

Hi,

Is there a way to extract rules from a trained a Artificial Neural Network using Weka?

If there isn't, how can I develop this feature and share with the main weka package?

Att,
Marcelo Damasceno
Mestrando em Ciência da Computação - UECE
Professor do IFRN/Macau
Fortaleza - Ceará - Brasil
+55 85 8797-9153
+55 84 96316990
+55 85 3341-0305

"Computers are like air conditioners: they stop working properly when you open windows". - EFR
--

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

How to evaluate a new boosting algorithm

2009-11-25T04:40:16Z

Hi all,

I've finished the implementation of a new Boosting algorithm for classification problem using the regression REPTree as weak leaner. To evaluate the new classifier, I deleted all class values of the training data set to create the testing data set, then I used the class Evaluation to evaluate my classifier with the new testing data set. In comparison with the training data, I can see that all instances of the testing data set are well classified (100% of instances is well classified). Is this enough to confirm that my classifier is good ? Is there any other method to evaluate a new classifier ?

Thanks for your help
Cheers,
VDung

--- On Wed, 11/25/09, Peter Reutemann <fracpete@...> wrote:

From: Peter Reutemann <fracpete@...>
Subject: Re: [Wekalist] Iterating through instances
To: "Weka machine learning workbench list." <wekalist@...>
Date: Wednesday, November 25, 2009, 3:35 PM

> I'm attempting write a clusterer for Weka.
> I'm having trouble getting the individual attribute values for a data
> instance. I'm handling numeric and nominal types.
>
> In going through the Weka code and online documentation, I'm confused
> on how to get an individual value from the data instances.

What kind of value do you want to have? Internal? String representation?

"value(Attribute/int)" returns the actual value for numeric attributes
and for nominal/string/relational the index. Copy/paste from the
Javadoc:
"the specified value as a double (If the corresponding attribute is
nominal (or a string) then it returns the value's index as a double)."

If you want to retrieve the actual "string" representation of
non-numeric attributes (that includes Date attributes!), use
"stringValue(Attribute/int)". Copy/paste from the Javadoc:
"Returns the value of a nominal, string, date, or relational attribute
for the instance as a string. "

Most algorithms just use the index of nominal/string/relational
attributes internal and not the string representation (to save space),
i.e., using the internal format retrieved via "value(Attribute/int)".

> See code
> below:
>
>
> int numAttrs = data.numAttributes();
> int numInstances = data.numInstances();
>
> for (int instIdx = 0; instIdx < numInstances; instIdx++) {
> Instance currInst = data.instance(instIdx);
>
> for (int attrIdx = 0; attrIdx < numAttrs; attrIdx++) {
> Attribute currAttr = currInst.attribute(attrIdx);
> if (currAttr.isNominal()) {
> // now what?
> }
> else if (currAttr.isNumeric()) {
> // now what?
> }
> }
> }

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Iterating through instances

2009-11-25T00:35:09Z

> I'm attempting write a clusterer for Weka.
> I'm having trouble getting the individual attribute values for a data
> instance. I'm handling numeric and nominal types.
>
> In going through the Weka code and online documentation, I'm confused
> on how to get an individual value from the data instances.

What kind of value do you want to have? Internal? String representation?

"value(Attribute/int)" returns the actual value for numeric attributes
and for nominal/string/relational the index. Copy/paste from the
Javadoc:
"the specified value as a double (If the corresponding attribute is
nominal (or a string) then it returns the value's index as a double)."

If you want to retrieve the actual "string" representation of
non-numeric attributes (that includes Date attributes!), use
"stringValue(Attribute/int)". Copy/paste from the Javadoc:
"Returns the value of a nominal, string, date, or relational attribute
for the instance as a string. "

Most algorithms just use the index of nominal/string/relational
attributes internal and not the string representation (to save space),
i.e., using the internal format retrieved via "value(Attribute/int)".

> See code
> below:
>
>
> int numAttrs = data.numAttributes();
> int numInstances = data.numInstances();
>
> for (int instIdx = 0; instIdx < numInstances; instIdx++) {
> Instance currInst = data.instance(instIdx);
>
> for (int attrIdx = 0; attrIdx < numAttrs; attrIdx++) {
> Attribute currAttr = currInst.attribute(attrIdx);
> if (currAttr.isNominal()) {
> // now what?
> }
> else if (currAttr.isNumeric()) {
> // now what?
> }
> }
> }

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Model serialization and prediction

2009-11-25T00:29:28Z

> Ok, so what are the steps involved.
> Currently:
> Classifiers.buildClassifier

No buildClassifier before cross-validating!

> Evaluation.crossValidateModel
> -> Save Model

crossValidateModel doesn't return a model (it builds as many as there
are folds and discards them after evaluating them).

> -> Save statistics
>
> If I don't have to give the trained model to the Eval, should I proceed that way ?
> Evaluation.crossValidateModel
> -> Save statistics
> Classifiers.buildClassifier
> -> Save Model

Correct. Of course, the model that you serialize is most likely
different from the other x (x = number of folds) during CV, as it is
generated from the full dataset.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

RE: Model serialization and prediction

2009-11-25T00:02:39Z

-----Message d'origine-----
De : wekalist-bounces@... [mailto:wekalist-bounces@...] De la part de Peter Reutemann
Envoyé : mardi 24 novembre 2009 19:03
À : Weka machine learning workbench list.
Objet : Re: [Wekalist] Model serialization and prediction

> Ok, got it. You have first to build the classifier, then to evaluate
> with crossValidateModel, and then you can serialize it...

The cross-validate is not necessary, only the call to buildClassifier(Instances).

NB: You shouldn't give the Evaluation class a trained model to cross-validate anyway. If the classifier doesn't reset itself completely, when its buildClassifier method is called, your results might be worthless ("information leakage").

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

Hi Peter.
Ok, so what are the steps involved.
Currently:
Classifiers.buildClassifier
Evaluation.crossValidateModel
-> Save Model
-> Save statistics

If I don't have to give the trained model to the Eval, should I proceed that way ?
Evaluation.crossValidateModel
-> Save statistics
Classifiers.buildClassifier
-> Save Model

Thanks
Fabien

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Iterating through instances

2009-11-24T21:34:20Z

I'm attempting write a clusterer for Weka.
I'm having trouble getting the individual attribute values for a data
instance. I'm handling numeric and nominal types.

In going through the Weka code and online documentation, I'm confused
on how to get an individual value from the data instances. See code
below:

int numAttrs = data.numAttributes();
int numInstances = data.numInstances();

for (int instIdx = 0; instIdx < numInstances; instIdx++) {
Instance currInst = data.instance(instIdx);

for (int attrIdx = 0; attrIdx < numAttrs; attrIdx++) {
Attribute currAttr = currInst.attribute(attrIdx);
if (currAttr.isNominal()) {
// now what?
}
else if (currAttr.isNumeric()) {
// now what?
}
}
}

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: weka classifiers

2009-11-24T18:00:34Z

> Can someone please tell me based on which features, weka classifiers perform
> the classification? Do i determine specific feature/s first or I should
> write something in the dataset or put the features in specific order?

On the commandline, you define via the -c parameter which attribute is
to be the class attribute. In the Explorer, you select the appropriate
attribute from the combobox on the preprocess panel.
The classifier will evaluate (what exactly happens here depends on the
actual classifier) all other attributes against this class attribute
then.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

weka classifiers

2009-11-24T17:52:54Z

Hi All,

Can someone please tell me based on which features, weka classifiers perform the classification? Do i determine specific feature/s first or I should write something in the dataset or put the features in specific order?

Many thanks,

Nancy

Windows Live: Friends get your Flickr, Yelp, and Digg updates when they e-mail you.
_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: addCluster

2009-11-24T12:46:15Z

>> is there any other way to display K-means cluster assignments to each
>> instance besides using AddCluster filter?
>
> Right-click in the results history list in the cluster panel of the
> Explorer, then choose "visualize cluster assignments".
>
> Thank you Peer, but this way show me the assignments in the graphical form.

You didn't specify in what form you wanted to "display" them.

> Some time ago, if I'm not wrong I found some thread about how to display the
> instances and its clustes assignmments with CLI but I'm not finding it any
> more, or I'm just confused. -)

Have you tried "-h" to list all the options for a clusterer? Then you
would come across the "-p" option...

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

RE: addCluster

2009-11-24T12:42:25Z

-----Original Message-----
From: wekalist-bounces@...
[mailto:wekalist-bounces@...] On Behalf Of Peter
Reutemann
Sent: Tuesday, November 24, 2009 11:47 AM
To: Weka machine learning workbench list.
Subject: Re: [Wekalist] addCluster

> is there any other way to display K-means cluster assignments to each
> instance besides using AddCluster filter?

Right-click in the results history list in the cluster panel of the
Explorer, then choose "visualize cluster assignments".

Cheers, Peter
....................

Thank you Peer, but this way show me the assignments in the graphical form.

Some time ago, if I'm not wrong I found some thread about how to display the
instances and its clustes assignmments with CLI but I'm not finding it any
more, or I'm just confused. -)

jcarlos

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Calculation of Information gain in weka.

2009-11-24T11:55:43Z

You would need to compute the entropy and/or gain for this on a per class basis
yourself.

Cheers,
Mark.

Hi Mark,

Any idea how? is it possible to send me one line of code

cheers,

Feras

--- On Tue, 11/24/09, Mark Hall <mhall@...> wrote:

From: Mark Hall <mhall@...>
Subject: Re: [Wekalist] Calculation of Information gain in weka.
To: "Weka machine learning workbench list." <wekalist@...>
Date: Tuesday, November 24, 2009, 3:38 PM

On 25/11/09 4:46 AM, Feras wrote:
>
> Hi Peter,
> I was able to get the information gain for each attribute in Weka, Thanks.
> Only one question remains:
> Is there a possibilty to get the weight for each attribute in each class?
> Info gain get the weight or the improtance for attribute in the whole
> dataset, but what about if we need to know the importance or weight for
> each attribute in each class?

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Re: polynomial regression

2009-11-24T11:47:00Z

On 24/11/09 8:09 PM, adrian golding wrote:
> hi, can someone please tell me how to do polynomial regression in WEKA?
> Do i use SMOreg?
> i am not going to do classification, i am more interested to find out
> the correlation coefficient of my dataset (of 2 attributes)

Since you've only got two attributes in your data, one simple approach would be
to use Weka's MathExpression filter to create some new attributes that are
copies of the original ones but raised to the power of n (2, 3, 4 etc). Then you
can just apply standard linear regression to this transformed data. Since linear
regression has automatic variable selection built in, it will choose the most
predictive terms to use.

Cheers,
Mark.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Calculation of Information gain in weka.

2009-11-24T11:38:33Z

On 25/11/09 4:46 AM, Feras wrote:
>
> Hi Peter,
> I was able to get the information gain for each attribute in Weka, Thanks.
> Only one question remains:
> Is there a possibilty to get the weight for each attribute in each class?
> Info gain get the weight or the improtance for attribute in the whole
> dataset, but what about if we need to know the importance or weight for
> each attribute in each class?

You would need to compute the entropy and/or gain for this on a per class basis
yourself.

Cheers,
Mark.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: addCluster

2009-11-24T10:46:48Z

> is there any other way to display K-means cluster assignments to each
> instance besides using AddCluster filter?

Right-click in the results history list in the cluster panel of the
Explorer, then choose "visualize cluster assignments".

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Question for different model prediction values

2009-11-24T10:45:29Z

Please no top-posting, see mailing list etiquette why
(http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html).

> For the different prediction values,
> I can make sense. That's true it comes from lack of precision/round.
> So, I have tried to find how to extend significant decimal precision.
> I found a following post and I installed developer version.
>
> http://old.nabble.com/can-the-precision-of-the-prediction-results-be-modified--tp23799213p23840294.html
>
> However, I can not find the number of output decimals option.
> Could please let me know about it?

Quote from the old post you mentioned:
"I've just added that to the output generators and committed a fix to
subversion."

Did you follow the instructions in the FAQ mentioned in that post?

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

addCluster

2009-11-24T10:41:53Z

Hi,

is there any other way to display K-means cluster assignments to each instance besides using AddCluster filter?

thanks,

jCarlos

Re: Question for different model prediction values

2009-11-24T10:41:09Z

Thanks Peter,

For the different prediction values,
I can make sense. That's true it comes from lack of precision/round.
So, I have tried to find how to extend significant decimal precision.
I found a following post and I installed developer version.

http://old.nabble.com/can-the-precision-of-the-prediction-results-be-modified--tp23799213p23840294.html

However, the coefficient of model does not be changed.
Is there any way to extend the precision for the model's coefficients?
How can I get a model that has more precision of the coefficient?

I really appreciate your help.

bshan

Peter Reutemann-3 wrote:

> I am wondering when I compared the prediction results between weka's
> prediction and direct calculation with its model.
> I used linear regression and M5Rules models.
>
> the test input data as followings:
>
> x1, x2
> -128.915, -0.1719
> -59.892, -0.0424
> -67.62, -0.01
>
> weka's Classifier model as
> pred = -0.0004 * x1 + 0.8953 * x2 + -0.0444

NB: These numbers are rounded.

>
> the prediction results when I use weka's prediction function are
> === Predictions on test set ===
>
> inst#, actual, predicted, error
> 1 -0.73 -0.144 0.586
> 2 0.63 -0.057 -0.687
> 3 -0.01 -0.025 -0.015
>
> But, when I calculate the prediction values using the weka's Classifier
> model directly it is different as followings:
>
> 1) -0.1467

Pretty close to -0.144

> 2) -0.0584

Also pretty close to -0.057

> 3) 0.0181

A bit further off -0.025

> So, I am wondering what's wrong? Why these are different?

Probably the usual rounding issues when displaying the model. But
without access to the original data (and how you calculated your
result), it's hard to tell.

> The other question is how can I use the result(rule) of M5Rules classifier
> model.
> In the result, it has some conditions (rules).
> I am confusing how can I use the result to calculate prediction value in my
> program directly.
> If you give me an example, it would be better to understand for me.

If you want to use a trained classifier, you might want to read FAQ
"How do I use WEKA's classes in my own code?". Link to the FAQs
available from the Weka homepage.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@list.scms.waikato.ac.nz
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html

Re: Calculation of Information gain in weka.

2009-11-24T10:06:01Z

> I was able to get the information gain for each attribute in Weka, Thanks.

OK.

> Only one question remains:
> Is there a possibilty to get the weight for each attribute in each class?
>
> Info gain get the weight or the improtance for attribute in the whole dataset, but what about if we need to know the importance or weight for each attribute in each class?

No idea.

Cheers, Peter
--
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/ Ph. +64 (7) 858-5174

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@...
List info and subscription status: https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html