Nabble - WEKA

Re: Feature selection

2008-10-11T02:04:50Z

Association mining is well-documented in weka, try the explorer or the
java api

pre-processing: for text use the filters like stringtowordvector or
nominaltostring (i don't remember the name of this one correctly)
Feature selection: for text i usually use Latent semantic analysis with
ranker search

for non-text: i normally use attributeSelection, which allows you to
specify which attribute evaluation method and which search method to use
when selecting/searching for the best attributes
the weka wiki also discusses these issues, just google weka wiki, and
search the wiki for these things

ashwin

svpriyan wrote:

> I am actually doing an experimental study with a few WEKA
> classification/clustering/ association analysis techniques on a few UCI (or
> other) datasets.
> I know how to do Preprocessing in WEKA, also I got the idea of comparing to
> classifiers
> My problems are
> • How I can do Association analysis for my data set
> { if I get some step by step hints I can try}
> • Further preprocessing
> I have to look on some of these techniques too in WEKA
> – Vector space for clustering/classification?
> – Feature selection/transformation
> How I can do these for my preprocessed data.
>
> Vass
>
>
>
>
>
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> Wekalist mailing list
> Wekalist@...
> https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
>

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Feature selection

2008-10-11T00:55:46Z

Hi, there are plenty of such techniques (association analysis, feature
selection etc) into WEKA.
I guess you can learn how to use them if you dedicate 1 hour of your
time to read the WEKA user manual.

regards

On Sat, Oct 11, 2008 at 9:50 AM, svpriyan <svpriyan@...> wrote:

>
>
> I am actually doing an experimental study with a few WEKA
> classification/clustering/ association analysis techniques on a few UCI (or
> other) datasets.
> I know how to do Preprocessing in WEKA, also I got the idea of comparing to
> classifiers
> My problems are
> • How I can do Association analysis for my data set
> { if I get some step by step hints I can try}
> • Further preprocessing
> I have to look on some of these techniques too in WEKA
> – Vector space for clustering/classification?
> – Feature selection/transformation
> How I can do these for my preprocessed data.
>
> Vass

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Feature selection

2008-10-11T00:50:57Z

I am actually doing an experimental study with a few WEKA classification/clustering/ association analysis techniques on a few UCI (or other) datasets.
I know how to do Preprocessing in WEKA, also I got the idea of comparing to classifiers
My problems are
• How I can do Association analysis for my data set
{ if I get some step by step hints I can try}
• Further preprocessing
I have to look on some of these techniques too in WEKA
– Vector space for clustering/classification?
– Feature selection/transformation
How I can do these for my preprocessed data.

Vass

RBF network width

2008-10-10T16:54:38Z

How does weka calculate the width of a hidden node in the rbf network?

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Feature selection

2008-10-10T16:50:03Z

> Further preprocessing
>
> – Vector space for clustering/classification?
> – Feature selection/transformation
> Could you please give some hints to do this.

Could you please elaborate on this?

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Could anyone tell me

2008-10-10T16:48:55Z

> I'm using WEKA classifiers to classify my data. I was using preset 10 fold
> cross validation and getting ROC values ranging from 0.545 to 0.773. Today I
> accidentally clicked Percentage Split 66 and ROC values for some of the
> classifiers went upto 0.864. Could anyone be able to explain this to me and
> is there anyway to improve the ROC values. I would appreciate that.

Classifiers are quite often sensitive to the order the data presented.
A single run of percentage split will produce most likely results that
are either too pessimistic or too optimistic. 10-fold CV gives you a
better estimate. Better even, the approach that the Weka Experimenter
uses, is 10 runs of 10-fold CV.

BTW the Explorer is only a tool for "exploring" the data, having a
play with it, getting a feel for it. If you want reliable statistical
results, use the Experimenter.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Could anyone tell me

2008-10-10T16:00:49Z

Hi,

I'm using WEKA classifiers to classify my data. I was using preset 10 fold cross validation and getting ROC values ranging from 0.545 to 0.773. Today I accidentally clicked Percentage Split 66 and ROC values for some of the classifiers went upto 0.864. Could anyone be able to explain this to me and is there anyway to improve the ROC values. I would appreciate that.

Thanks!

-Mehdi

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Feature selection

2008-10-10T01:12:21Z

Dear Sir,

Further preprocessing

– Vector space for clustering/classification?
– Feature selection/transformation
Could you please give some hints to do this.

Vass

Feature selection

2008-10-09T23:49:14Z

Dear Sir,

Further preprocessing

– Vector space for clustering/classification?
– Feature selection/transformation
Could you please give some hints to do this.

Vass

Re: predicted class in 10 fold cross validation

2008-10-09T11:57:24Z

> I have done 10 fold cross validation in Weka using LWL classifier on a data
> set with 30 instances. Is there a way to find out the predicted class for
> each of these 30 instances?

- In the Explorer: check "Output predictions" in the "More options..." dialog
- On the command-line (since 3.5.8): use the "-p" option (if you need
to do it for a numeric class, you need download a developer snapshot
from the Weka homepage, since the 3.5.8 release contains a bug -
didn't output anything)
- From Java: use method "predictions()" of the
weka.classifiers.Evaluation class (only nominal classes)

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Convert CSV 2 ARFF in Java code?

2008-10-09T11:50:21Z

> Thank you for your email, actually I need help on detailed steps of how to predict phosphorylation sites with my existing file.
> Would you please tell me how to do that?

See FAQ "How do I use Weka's classes in my own code?" for more
information of how to use Weka's API from your 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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: java code for predicting contineous values using MLP-neural network

2008-10-09T11:47:24Z

> I was to get the predicted values using java code if i am having predictable
> column as discrete
> but i am facing a problem if the preditable column is contineous............
> i am trying the code for multilayer perceptron.............
>
> can any one give me a sample java code for prediction for multilayer
> perceptron for contineous values

You just use the classifyInstance method, like in the example here:
http://weka.sourceforge.net/wiki/index.php/Use_Weka_in_your_Java_code#Classifying_instances

This method returns, in case of numeric attributes, the numeric prediction.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

RE: Convert CSV 2 ARFF in Java code?

2008-10-09T09:39:22Z

Dear Peter,

Thank you for your email, actually I need help on detailed steps of how to predict phosphorylation sites with my existing file.
Would you please tell me how to do that?

Thanks,
Bahareh

-----Original Message-----
From: wekalist-bounces@... on behalf of Peter Reutemann
Sent: Mon 06/10/2008 4:22 PM
To: Weka machine learning workbench list.
Subject: Re: [Wekalist] Convert CSV 2 ARFF in Java code?

> I am a new WEKA user and I don't know how to convert my CSV file to arff. I
> use eclipse for java programming but it doesn't recognize weka. Would you
> please help me to find out where I have to type this code?

[...]

Steps involved:
- Create a new project (or use an existing one)
- add the weka.jar as (external) jar under "Libraries"
(or, in an existing projects, under "Project -> Properties -> Java
Build Path -> Libraries")
- create a new package, e.g., "myweka"
- add the CSV2Arff class in that package and set the package of the
source file to "myweka"

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

winmail.dat (4K) Download Attachment

predicted class in 10 fold cross validation

2008-10-09T08:44:06Z

Hi all,

I have done 10 fold cross validation in Weka using LWL classifier on a data set with 30 instances. Is there a way to find out the predicted class for each of these 30 instances?

Thanks
Rk

Add more friends to your messenger and enjoy! Invite them now.
_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

RE: Convert CSV 2 ARFF in Java code?

2008-10-09T08:22:28Z

Dear Peter,

Thank you for your email, actually I need help on detailed steps of how to predict phosphorylation sites with my existing file.

Thanks,
Bahareh

-----Original Message-----
From: wekalist-bounces@... on behalf of Peter Reutemann
Sent: Mon 06/10/2008 4:22 PM
To: Weka machine learning workbench list.
Subject: Re: [Wekalist] Convert CSV 2 ARFF in Java code?

> I am a new WEKA user and I don't know how to convert my CSV file to arff. I
> use eclipse for java programming but it doesn't recognize weka. Would you
> please help me to find out where I have to type this code?

[...]

Steps involved:
- Create a new project (or use an existing one)
- add the weka.jar as (external) jar under "Libraries"
(or, in an existing projects, under "Project -> Properties -> Java
Build Path -> Libraries")
- create a new package, e.g., "myweka"
- add the CSV2Arff class in that package and set the package of the
source file to "myweka"

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

winmail.dat (4K) Download Attachment

java code for predicting contineous values using MLP-neural network

2008-10-09T02:45:16Z

hi,

I am new to WEKA,I was able to get the predicted values using java code if i am having predictable column as discrete
but i am facing a problem if the preditable column is contineous............
i am trying the code for multilayer perceptron.............

can any one give me a sample java code for prediction for multilayer perceptron for contineous values

Re: ROC for cross validation

2008-10-09T02:36:45Z

> So it's not a very sophisticated way to make a ROC for cross validations -
> there is no averaging being done as is the case for other methods like
> Vertical Averaging or Threshold Averaging for instance.
>
>
> There's no need for averaging with just 1 run of CV.
>
>
> Hmm, I do not agree here. You could still build the average of the curves of
> all folds to see how "stable" the curves are located in their space. Here
> is, for example, how this is done by RapidMiner for a cross validation of
> one learning scheme (in red are the ROC values and in blue are the
> thresholds):
>
> http://www.rapid-i.com/roc/RapidMiner_ROC_1.jpg
>
> This gives you an idea how robust a threshold selection for example in
> cost-sensitive learning is. And here is another result, this time of a
> 10-fold cross validation of three different learning schemes (generated with
> the RapidMiner operator "ROCComparison"):
>
> http://www.rapid-i.com/roc/RapidMiner_ROC_2.jpg
>
> Maybe these pictures give you an idea why something like this can actually
> help.

True, does anybody feel like contributing some code? ;-)

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: ROC for cross validation

2008-10-09T02:33:48Z

Hi,

So it's not a very sophisticated way to make a ROC for cross validations -
there is no averaging being done as is the case for other methods like
Vertical Averaging or Threshold Averaging for instance.


There's no need for averaging with just 1 run of CV.

Hmm, I do not agree here. You could still build the average of the curves of all folds to see how "stable" the curves are located in their space. Here is, for example, how this is done by RapidMiner for a cross validation of one learning scheme (in red are the ROC values and in blue are the thresholds):

http://www.rapid-i.com/roc/RapidMiner_ROC_1.jpg

This gives you an idea how robust a threshold selection for example in cost-sensitive learning is. And here is another result, this time of a 10-fold cross validation of three different learning schemes (generated with the RapidMiner operator "ROCComparison"):

http://www.rapid-i.com/roc/RapidMiner_ROC_2.jpg

Maybe these pictures give you an idea why something like this can actually help.

Cheers,
Ingo

-- 
Ingo Mierswa
Managing Director

Rapid-I GmbH
Stockumer Str. 475
44149 Dortmund, Germany

Phone: +49 (0)231 425 786 90

E-Mail:  mierswa@...

Sitz: Dortmund
HRB 20720, Amtsgericht Dortmund
Geschäftsführer: Ingo Mierswa, Ralf Klinkenberg

www: http://rapid-i.com/

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: ROC for cross validation

2008-10-09T00:38:58Z

> I've looked at the ThresholdCurve code and I wanted to confirm that when k
> fold cross validation is performed, the curve that is produced is simply a
> ROC curve made from all the predictions from the k folds.

Yes, correct.

> So it's not a very sophisticated way to make a ROC for cross validations -
> there is no averaging being done as is the case for other methods like
> Vertical Averaging or Threshold Averaging for instance.

There's no need for averaging with just 1 run of CV.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

ROC for cross validation

2008-10-08T22:41:37Z

Hello,
I've looked at the ThresholdCurve code and I wanted to confirm that
when k fold cross validation is performed, the curve that is produced
is simply a ROC curve made from all the predictions from the k folds.
So it's not a very sophisticated way to make a ROC for cross
validations - there is no averaging being done as is the case for
other methods like Vertical Averaging or Threshold Averaging for
instance.
thanks,
shirley

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Weka moves from CVS to Subversion

2008-10-08T19:05:51Z

Hi folks,

Weka's CVS repository has been migrated to Subversion. You can access
it from:

https://svn.scms.waikato.ac.nz/svn/weka/

At present the "trunk" directory has anonymous read access. We are
just waiting on our technical support guys to add anonymous read
access to "branches" and "tags" as well.

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 7 847-3537 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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T14:56:37Z

>> Well, I call it "unreliable" since you ask for "1" neighbor, but you
>> can end up with more than that (in my tests, I quite often got 2
>> neighbors back). IMHO it's misleading having the method called
>> "kNearestNeighbours"; it should be called "kOrMoreNearestNeighbors".
>
> How would you break ties then? If there are more than one nearest neighbour
> with the same distance, a more reliable model will be generated by using all
> of them. IBk has always done this - when you set k to 1 and there are ties,
> it returns the majority class (or, more correctly, generates a distribution
> from the votes for each class) over these ties as the prediction.

Good point.

But what I'm saying is that the method is misleading (as well as the
Javadoc documentation), it violates design by contract. Other schemes
expect to get exactly k neighbors back and therefore don't check if
they get back more and need to break ties themselves. One would have
to check whether the current implementation doesn't break any schemes,
due to not checking the actual returned number of instances.

At least there should be an option available for the user to choose
truncating the returned instances to the requested number, for
instance "tieBreaking" property (command-line option "-T <type>"):
- none, include ties in result
- first come, first served
- pick majority class

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T14:39:46Z

On 9/10/2008, at 10:07 AM, Peter Reutemann wrote:

>>> The LinearNNSearch algorithm (package weka.core.neighboursearch -
>>> Weka
>>> 3.5.8) is a bit unreliable in case you have ties in distances. As
>>> far
>>> as I know, if the algorithm encounters ties on the last neighbor
>>> to be
>>> returned, it will return all of them. I.e., if you want 10 neighbors
>>
>> This is intentional and not "unreliable". In fact, it returns ties
>> at any
>> instance in the top k, not only the last one. So, if you ask for five
>> nearest neighbours and the closest neighbour has two ties, then
>> three of the
>> top five slots are consumed.
>
> Well, I call it "unreliable" since you ask for "1" neighbor, but you
> can end up with more than that (in my tests, I quite often got 2
> neighbors back). IMHO it's misleading having the method called
> "kNearestNeighbours"; it should be called "kOrMoreNearestNeighbors".

How would you break ties then? If there are more than one nearest
neighbour with the same distance, a more reliable model will be
generated by using all of them. IBk has always done this - when you
set k to 1 and there are ties, it returns the majority class (or,
more correctly, generates a distribution from the votes for each
class) over these ties as the prediction.

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 7 847-3537 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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T14:07:27Z

>> The LinearNNSearch algorithm (package weka.core.neighboursearch - Weka
>> 3.5.8) is a bit unreliable in case you have ties in distances. As far
>> as I know, if the algorithm encounters ties on the last neighbor to be
>> returned, it will return all of them. I.e., if you want 10 neighbors
>
> This is intentional and not "unreliable". In fact, it returns ties at any
> instance in the top k, not only the last one. So, if you ask for five
> nearest neighbours and the closest neighbour has two ties, then three of the
> top five slots are consumed.

Well, I call it "unreliable" since you ask for "1" neighbor, but you
can end up with more than that (in my tests, I quite often got 2
neighbors back). IMHO it's misleading having the method called
"kNearestNeighbours"; it should be called "kOrMoreNearestNeighbors".

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T13:58:40Z

On 9/10/2008, at 8:38 AM, Peter Reutemann wrote:

>>
>
> The LinearNNSearch algorithm (package weka.core.neighboursearch - Weka
> 3.5.8) is a bit unreliable in case you have ties in distances. As far
> as I know, if the algorithm encounters ties on the last neighbor to be
> returned, it will return all of them. I.e., if you want 10 neighbors

This is intentional and not "unreliable". In fact, it returns ties at
any instance in the top k, not only the last one. So, if you ask for
five nearest neighbours and the closest neighbour has two ties, then
three of the top five slots are consumed.

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 7 847-3537 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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T12:38:34Z

> I have still another question about how LWL works.
> Assume I use LWL with naive Bayes.
>
> If I set k=total number of instances in the training set and if I
> choose the constant kernel, then LWL is actually equivalent to a
> global naive Bayes.
> Ok.
>
> However, let me assume k=1.
> Now, the LWL should be equivalent to a Naive Bayes built on the
> closest instance.
> This is not the case, however.
>
> Let me consider the first instance of the below data set. If I build a
> Naive Bayes using this only instance as training set,
> the instance itself is classified by naive Bayes as 0.214 0.214 *0.57.
>
> However, If I run LWL with naive Bayes and a k=1
> (LWL -U 5 -K 1 -W ...naiveBayes), I get the following classification:
> 0.25 0.25 *0.5.
>
> Why do the results differ?
>
> Moreover, working with k=1 I get different results, if I use different
> kernels. I expected instead the results to be insensitive on the
> choice of the kernel for k=1, as in fact there is only one instance
> under consideration, to which all kernels should give weight 1.
>
> What do I miss?

The LinearNNSearch algorithm (package weka.core.neighboursearch - Weka
3.5.8) is a bit unreliable in case you have ties in distances. As far
as I know, if the algorithm encounters ties on the last neighbor to be
returned, it will return all of them. I.e., if you want 10 neighbors
and you encounter ties with several instances, say 3 have the same
distance, on the tenth one, then all of these will be returned, in
other words 12. I just ran LWL with NaiveBayes and CONSTANT weighting
(-U 5) and k=1 on the UCI dataset "anneal" and encountered sometimes
that 2 instead of 1 instance were returned and use as training set for
the base classifier. That might explain your observations.

I leave that for the Weka maintainer...

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-08T09:07:05Z

I have still another question about how LWL works.
Assume I use LWL with naive Bayes.

If I set k=total number of instances in the training set and if I
choose the constant kernel, then LWL is actually equivalent to a
global naive Bayes.
Ok.

However, let me assume k=1.
Now, the LWL should be equivalent to a Naive Bayes built on the
closest instance.
This is not the case, however.

Let me consider the first instance of the below data set. If I build a
Naive Bayes using this only instance as training set,
the instance itself is classified by naive Bayes as 0.214 0.214 *0.57.

However, If I run LWL with naive Bayes and a k=1
(LWL -U 5 -K 1 -W ...naiveBayes), I get the following classification:
0.25 0.25 *0.5.

Why do the results differ?

Moreover, working with k=1 I get different results, if I use different
kernels. I expected instead the results to be insensitive on the
choice of the kernel for k=1, as in fact there is only one instance
under consideration, to which all kernels should give weight 1.

What do I miss?

Thank you very much

@attribute tear-prod-rate {reduced,normal}
@attribute contact-lenses {soft,hard,none}
@data
reduced,none
normal,soft
normal,hard

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: About CV & confusion matrix

2008-10-08T02:51:16Z

> I have a question about how weka calculates confusion matrix?
> There suppose to be Y models and Y confusion matrixes for Y folds CV.
> However, there are only one confusion matrix printed by weka.
> How does the weka calculated such confusion matrix?

In the course of one cross-validation, the full dataset will be used
for testing (10-fold CV gives you 10 disjoint test sets that make up
the full dataset). Weka collects the information, like distributions
and (mis)classifications, for all the disjoint test sets and then
generates the output. Therefore it's only 1 confusion matrix that gets
generated and not 10.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

About CV & confusion matrix

2008-10-08T00:04:27Z

Dear sir:

I have a question about how weka calculates confusion matrix?
There suppose to be Y models and Y confusion matrixes for Y folds CV.
However, there are only one confusion matrix printed by weka.
How does the weka calculated such confusion matrix?
thanks

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: total loss precision M5P regression coefficients problem

2008-10-07T16:10:24Z

>>> To recap, the M5P classifier output displays the leaf regression models with
>>> a fixed float format such that any coefficient with an absolute value <
>>> .00005 will display as zero -- a total loss of precision error.
>>>
>>> It's been discussed that this is not an internal representation problem,
>>> merely a output formatting string limitation. (Is it the case that Java
>>> lacks a smart formatting routine for floats analogous, say, to %g in C?)
>>>
>
> Java has a similar facility, e.g. System.out.format("x = %g", x);
> IMHO it would be great if "someone" could go through Weka
> and consistently replace printing of doubles with the above,
> but I do understand that this is a lot of tedious and boring work.

Mostly it would be replacing the Utils.doubleToString methods being
used. There are about 500 occurrences in the current developer version
of Weka. But still, with an IDE like Eclipse it is fairly easy
locating them, ascertain whether it should be replaced (one could
limit it to the toString() methods of classifiers, etc.) and replace
them if necessary.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: total loss precision M5P regression coefficients problem

2008-10-07T15:57:57Z

Hi

>> To recap, the M5P classifier output displays the leaf regression models with
>> a fixed float format such that any coefficient with an absolute value <
>> .00005 will display as zero -- a total loss of precision error.
>>
>> It's been discussed that this is not an internal representation problem,
>> merely a output formatting string limitation. (Is it the case that Java
>> lacks a smart formatting routine for floats analogous, say, to %g in C?)
>>

Java has a similar facility, e.g. System.out.format("x = %g", x);
IMHO it would be great if "someone" could go through Weka
and consistently replace printing of doubles with the above,
but I do understand that this is a lot of tedious and boring work.

Bernhard

---------------------------------------------------------------------
Bernhard Pfahringer, Dept. of Computer Science, University of Waikato
http://www.cs.waikato.ac.nz/~bernhard +64 7 838 4041

_______________________________________________
Wekalist mailing list
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Replace missing values using the EM algorithm

2008-10-07T14:27:05Z

A PhD student (and Google summer of code contributer) is currently
working on a filter to perform Bayesian multiple imputation for
missing values (as described in Schafer, J. L. Analysis of Incomplete
Multivariate Data, New York: Chapman and Hall, 1997). Hopefully he
will be contributing this in the near future.

Cheers,
Mark.

On 2/10/2008, at 9:42 PM, ali Aich wrote:

> Hi,
>
> Now I work with the bayes net algorithms. For replacing missing
> values, I use the "MissingValueFilter" of Weka. I want to know how
> we could use the algorithm EM in order to replace missing values.
>
> Thank's for your responses.
> Cordially
>
> Votre correspondant a choisi Hotmail et profite d'un stockage
> quasiment illimité. Créez un compte Hotmail gratuitement !
> _______________________________________________
> Wekalist mailing list
> Wekalist@...
> https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

--
Mark Hall
Senior Developer/Consultant, Pentaho Open Source Business Intelligence
Citadel International, Suite 340, 5950 Hazeltine National Dr.,
Orlando, FL 32822, USA
+64 7 847-3537 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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: SMO Probabilistic

2008-10-07T13:37:42Z

Turning on probabilistic prediction for SMO results in a standard SMO
model being built and then a logistic regression fitted to the output
- i.e. a new data set is created that contains the predicted
responses of SMO along with the original class values. A logistic
regression is then fitted to this derived data set in order to
produce probability estimates.

You can see the probabilities produced by turning on the option to
output probabilities in the Explorer or by using the -p option on the
command line.

Cheers,
Mark.

On 8/10/2008, at 1:44 AM, Thomas Oommen wrote:

> I'm doing a two class SMO classification using the polynomial
> Kernel. I use
> 5-fold cross validation.
> When I turn the probabilistic prediction TRUE. I get the following
> output,
> other than the probabilities and accuracy.
>
> Logistic Regression with ridge parameter of 1.0E-8
> Coefficients...
> Class
> Variable Yes
> ====================
> pred -2.8806
> Intercept -1.4197
>
>
> Odds Ratios...
> Class
> Variable Yes
> ====================
> pred 0.0561
>
> Can anyone explain what this means?
>
> Thanks in adavance.
>
> Thomas Oommen
> Tufts University
> Medford, MA-02155
>
>
>
> _______________________________________________
> Wekalist mailing list
> Wekalist@...
> https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: total loss precision M5P regression coefficients problem

2008-10-07T13:37:08Z

> To recap, the M5P classifier output displays the leaf regression models with
> a fixed float format such that any coefficient with an absolute value <
> .00005 will display as zero -- a total loss of precision error.
>
> It's been discussed that this is not an internal representation problem,
> merely a output formatting string limitation. (Is it the case that Java
> lacks a smart formatting routine for floats analogous, say, to %g in C?)
>
> Anyway, I'm come up against this problem, and it's causing me enough grief
> to submit this question. The obvious work around is to normalise the input
> values in the data sets so that all the input values are in ranges that will
> increase the coefficients to visible levels, but this introduces a whole
> range of additional problems with respect to managing multiple versions of
> data sets, scope for errors and inconsistency, etc. In short, it would be
> far preferable to be able to avoid messing around with all that.
>
> So, two questions: a) Is there a patch or version of Weka (I'm using 3.4.13)
> that has a smarter print formatter for floats?

Not that I know of. Most people don't seem to have a problem with that.

> b) If not, is there a good
> technical reason why this can't be/should not be fixed once and for all?

No, just they way it was implemented (probably not to waste too much
space). You can always change the output yourself. Just modify method
"treeToString(int)" of class weka.classifiers.trees.m5p.RuleNode.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: how to inspect the weights of LWL

2008-10-07T13:33:41Z

> I would like to inspect the weights assigned by LWL to the different
> instances of the training set.
> (Of course, the weights will change every time a new classification is
> performed, so ).
>
> I have turned on the debug option in LWL, but this did not change the
> output I get.
> Is there a way to inspect the weights assigned to the different
> instances of the training set?

LWL is a lazy learning scheme, i.e., the actual work is done at
prediction time. Each time you call distributionForInstance or
classifyInstance to make a prediction, a new dataset (based on the
training set) gets generated with the new weights assigned. Check out
variable "neighbours" in the distributionForInstance method, which is
used to train the base classifier with. If you want to output some
information, you will have to hack the code accordingly.

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
Wekalist@...
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist