WEKA machine learning software discussion

Harri Saarikoski | 1 Jan 2010 05:39

Re: Attribute Ranking in Weka

2009/12/31 Feras <feras_o <at> yahoo.com>

Hi,

Is there a way in Weka to rank the attributes from best to worse by giving them numeric numbers presenteing their importance in data.

see manually insertable feature weights -> xrff format in weka docs
limited to some classifiers only -> see wekalist archives on feature/attribute weights

If available, please provide me with the code or link to do it using some sample dataset.

Cheers,

Feras

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

--
-----------------
Harri M.T. Saarikoski
M.A, PhD graduate student
Helsinki University
Finland

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Joanna Pschorr | 1 Jan 2010 18:02

Re: How to understand data format after applying StringToWordVector?

>> My Question is can anybody explain what the data means after applying
>> StringToWordVector? See  http://old.nabble.com/file/p26970215/data.txt
>> data.txt  data file. I know StringToWordVector transform data into sort of
>> TF-IDF transformation.
>> But see a test instance before and after applying the StringToWordVector:
>>
>>  <at> data
>> 'matricid stori hors wayn panti pop highwai noir thriller refer ',5
>>
>>   <at> data
>> {0 5,150 1.555245,201 1.662094,207 2.268923,219 2.623,232 1.606613,262
>> 0.793745,279 2.268923,305 2.423594}
>>
>> what 0 5, 150 1.555245 means?
>>
>
Hi,

I must correct myself: The pair of values 0 5 is the class of course,
I must have overlooked that your class names are numbers, sorry for
that. Terms not found in the training data can't possibly appear in
the vector or even get a weight. The attribute containing class names
(Trainraiting) is the first one (position 0) and the category of your
test data record is 5 so then it makes sense.

Regards,
\joanna

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

musi | 1 Jan 2010 18:48

Re: How to understand data format after applying StringToWordVector?


Yes, I got it it makes sense.
thanks,

Joanna Pschorr wrote:
> 
>>> My Question is can anybody explain what the data means after applying
>>> StringToWordVector? See  http://old.nabble.com/file/p26970215/data.txt
>>> data.txt  data file. I know StringToWordVector transform data into sort
>>> of
>>> TF-IDF transformation.
>>> But see a test instance before and after applying the
>>> StringToWordVector:
>>>
>>>  <at> data
>>> 'matricid stori hors wayn panti pop highwai noir thriller refer ',5
>>>
>>>   <at> data
>>> {0 5,150 1.555245,201 1.662094,207 2.268923,219 2.623,232 1.606613,262
>>> 0.793745,279 2.268923,305 2.423594}
>>>
>>> what 0 5, 150 1.555245 means?
>>>
>>
> Hi,
> 
> I must correct myself: The pair of values 0 5 is the class of course,
> I must have overlooked that your class names are numbers, sorry for
> that. Terms not found in the training data can't possibly appear in
> the vector or even get a weight. The attribute containing class names
> (Trainraiting) is the first one (position 0) and the category of your
> test data record is 5 so then it makes sense.
> 
> Regards,
> \joanna
> 
> _______________________________________________
> Wekalist mailing list
> Send posts to: Wekalist <at> 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
> 
> 

--

-- 
View this message in context: http://old.nabble.com/How-to-understand-data-format-after-applying-StringToWordVector--tp26970215p26986965.html
Sent from the WEKA mailing list archive at Nabble.com.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

musi | 1 Jan 2010 18:51

SMOreg error: exceptionjava.lang.ArrayIndexOutOfBoundsException:


I used SMOreg and I am getting the follwoing error while evaluating the
model (using the same dataset)

exceptionjava.lang.ArrayIndexOutOfBoundsException: 11
java.lang.ArrayIndexOutOfBoundsException: 11
        at weka.core.SparseInstance.index(Unknown Source)
        at
weka.classifiers.functions.supportVector.RegOptimizer.SVMOutput(Unknown
Source)
        at weka.classifiers.functions.SMOreg.classifyInstance(Unknown
Source)
        at
weka.classifiers.Evaluation.evaluateModelOnceAndRecordPrediction(Unknown
Source)
        at weka.classifiers.Evaluation.evaluateModel(Unknown Source)
        at
netflix.weka.SVMRegTextByAttribute.evaluatePredictiveModel(SVMRegTextByAttribute.java:617)
        at
netflix.weka.SVMRegTextByAttribute.doNBSteps(SVMRegTextByAttribute.java:338)
        at
netflix.weka.SVMRegTextByAttribute.main(SVMRegTextByAttribute.java:754)

Can someone explain why it is so? BTW I am using StringToWordVector with
"-T", "-O" options for converting text vectors into numeric values. My
output class in integer and lies in 1-5 range.

(I must mention, the same dataset works well for linearregression under
weka)
Thanks for your time, 
--

-- 
View this message in context: http://old.nabble.com/SMOreg-error%3A--exceptionjava.lang.ArrayIndexOutOfBoundsException%3A-tp26986984p26986984.html
Sent from the WEKA mailing list archive at Nabble.com.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Harri Saarikoski | 2 Jan 2010 12:31

Re: SMOreg error: exceptionjava.lang.ArrayIndexOutOfBoundsException:

2010/1/1 musi <eng.musi <at> gmail.com>

I used SMOreg and I am getting the follwoing error while evaluating the
model (using the same dataset)

exceptionjava.lang.ArrayIndexOutOfBoundsException: 11

while we wait for a weka code expert to give the list a hand, we may recycle peter's old replies from wekalist archives:
http://old.nabble.com/ArrayIndexOutOfBoundsException---due-to-change-in-training-dataset-td26602953.html

does that help you? has reference to StringToWordVector too...

java.lang.ArrayIndexOutOfBoundsException: 11
at weka.core.SparseInstance.index(Unknown Source)
at
weka.classifiers.functions.supportVector.RegOptimizer.SVMOutput(Unknown
Source)
at weka.classifiers.functions.SMOreg.classifyInstance(Unknown
Source)
at
weka.classifiers.Evaluation.evaluateModelOnceAndRecordPrediction(Unknown
Source)
at weka.classifiers.Evaluation.evaluateModel(Unknown Source)
at
netflix.weka.SVMRegTextByAttribute.evaluatePredictiveModel(SVMRegTextByAttribute.java:617)
at
netflix.weka.SVMRegTextByAttribute.doNBSteps(SVMRegTextByAttribute.java:338)
at
netflix.weka.SVMRegTextByAttribute.main(SVMRegTextByAttribute.java:754)

Can someone explain why it is so? BTW I am using StringToWordVector with
"-T", "-O" options for converting text vectors into numeric values. My
output class in integer and lies in 1-5 range.

(I must mention, the same dataset works well for linearregression under
weka)
Thanks for your time,
--
View this message in context: http://old.nabble.com/SMOreg-error%3A--exceptionjava.lang.ArrayIndexOutOfBoundsException%3A-tp26986984p26986984.html
Sent from the WEKA mailing list archive at Nabble.com.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

--
-----------------
Harri M.T. Saarikoski
M.A, PhD graduate student
Helsinki University
Finland

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Harri Saarikoski | 2 Jan 2010 12:44

Re: How to understand data format after applying StringToWordVector?

2009/12/30 musi <eng.musi <at> gmail.com>

Hi,

I am applying StringToWordVector over vectors of text. I have training set
and test set and I am using Java program to apply StringToWordVector and
training a classifier (Naibe Bayes, and SVM). I have 5 nominal classes
{1,2,3,4,5}.

My Question is can anybody explain what the data means after applying
StringToWordVector? See http://old.nabble.com/file/p26970215/data.txt
data.txt data file. I know StringToWordVector transform data into sort of
TF-IDF transformation.
But see a test instance before and after applying the StringToWordVector:

<at> data
'matricid stori hors wayn panti pop highwai noir thriller refer ',5

<at> data
{0 5,150 1.555245,201 1.662094,207 2.268923,219 2.623,232 1.606613,262
0.793745,279 2.268923,305 2.423594}

what 0 5, 150 1.555245 means?

this is weka arff in sparse format, convenient for a lot of binary attribute representation
it omits the attributes for each sample where value is zero, and thus saves time and space

in my word sense disambiguation set that I don't think is dissimilar to a text mining set:
<at> attribute 'keen<>to' {0,1}
<at> attribute 'pursued<>we' {0,1}
<at> attribute 'senseclass' {...}
<at> data
{13 1, 282 1, 493 1, 1033 1, 1388 1, 1712 1, 1791 1, 1961 1, 2528 1, 2884 1, 4308 1, 4704 1, 4900 art~1:06:00::}

first number is the attribute index number in the arff listing of attributes,
second is the only allowed non-zero value ('1') for all

I do think your second number has apparently a product of tf-idf transformation of columns
which originally (before transformation) were bunch of ones and zeroes
(i.e. attribute present or absent), apparently all the other attributes (e.g. 1-149)
have value zero (as in 'attribute has zero merit')?

Harri

I am applying the same StringToWordVector over training set and test set.
Furthermore, I am not getting good scores for this problem using Naive
Bayes. What can be wrong?

Thanks for your reply,
Rgrds,

Mustansar Ali Ghazanfar (Musi)
PhD, ISIS group, ECS,
University of Southampton, UK

--
View this message in context: http://old.nabble.com/How-to-understand-data-format-after-applying-StringToWordVector--tp26970215p26970215.html

Sent from the WEKA mailing list archive at Nabble.com.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

--
-----------------
Harri M.T. Saarikoski
M.A, PhD graduate student
Helsinki University
Finland

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

sunwing | 2 Jan 2010 12:51

Re: about REPtree


Dear Harri，

Thank you very much for your kind help.

Xiaoyu

--

-- 
View this message in context: http://old.nabble.com/about-REPtree-tp26964707p26992399.html
Sent from the WEKA mailing list archive at Nabble.com.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Nancy Adam | 2 Jan 2010 23:47

RMSE and R2

Hi everyone,

Happy New Year.

Can anyone please tell me in these questions?:

Is RMSE that weka computes based on the training set or based on the testing set?

Can R2 be computed from RMSE?

When do I say that a regression program is better than another one in terms of RMSE? i.e when its RMSE is bigger or smaller the other program?

Many thanks,

Nancy

Keep your friends updated— even when you’re not signed in.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Harri Saarikoski | 3 Jan 2010 11:16

Re: RMSE and R2

2010/1/3 Nancy Adam <nancyadam84 <at> hotmail.com>

Hi everyone,

Happy New Year.

Can anyone please tell me in these questions?:

Is RMSE that weka computes based on the training set or based on the testing set?

depends on the evaluation option:
if it's cross-validation you have selected, then on training set
if there's a separate test set provided, then on that test set

Can R2 be computed from RMSE?

same information is provided for both: actual outcomes and their predicted values -> offsets
so yes, they can : root mean squared error is square root of mean squared error
while R2 is the square of the sample correlation coefficient
-> critical difference of how they best apply to different cases
might be that R squared is bounded/normalised between -1 and 1
while RMSE 'modeling of error quantity' is unbounded 0 to infinity
(but I'm not sure)

When do I say that a regression program is better than another one in terms of RMSE? i.e when its RMSE is bigger or smaller the other program?

good = when error that RMSE measures is as small as poosible

Harri

Many thanks,

Nancy

Keep your friends updated— even when you’re not signed in.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

--
-----------------
Harri M.T. Saarikoski
M.A, PhD graduate student
Helsinki University
Finland

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

headers

Mark Hall | 3 Jan 2010 22:34

Re: about REPtree

Harri Saarikoski wrote:
> 
> 
> 2009/12/30 sunwing <ap.quxiaoyu <at> archi.kyoto-u.ac.jp 
> <mailto:ap.quxiaoyu <at> archi.kyoto-u.ac.jp>>
> 
> 
>     Dear all,
> 
>     1)I　ａｍ　ｗｏｎｄｅｒｉｎｇ　ｔｈｅ　ｍｅａｎｉｎｇ　ｏｆ　
>     ｎｕｍｂｅｒｓ　ｉｎ　ｔｈｅ　ｌｅａｖｅ　ｎｏｄｅ　ｏｆ　ａ　REPtree．
>     ｆｏｒ　ｅｘａｍｐｌｅ：
> 
>     ４:68.83 (41.57/1989.69) [24.01/4049.91]
>     I only know"4" at the beginning is the series number of node, but
>     what do
>     the following numbers mean?
> 
> 
> well, in a reptree built from cpu dataset:
> MMAX < 6100
> |   MMAX < 3500 : 20.03 (21/5.14) [13/5]
> |   MMAX >= 3500 : 28.5 (32/8.41) [12/23.49]
> MMAX >= 6100 : 40.89 (19/16.14) [14/41.38]
> 
> the bit before ":" is the conditional rule ("feature value <=> value 
> cutoff") that was built
> the first number after it is class value, to which the rule is attached
> which in sets with numeric class is naturally a number
> cf. reptree on nominal class sets (iris):
> petallength < 2.5 : Iris-setosa (33/0) [17/0]
> petallength >= 2.5
> |   petalwidth < 1.75 : Iris-versicolor (36/3) [18/2]
> |   petalwidth >= 1.75 : Iris-virginica (31/1) [15/0]
> 
> then come the brackets... phew
> 
> using reptree model on iris set as example:
> -> first figures in both brackets sum up to 150, so for each rule, 
> that's the /number /of instances matching that rule
> and their sum is total number of instances in the set
> -> latter figures in both brackets sum up to 6, so for each rule, this 
> is the /percentage /of instances misclassified by that rule
> and their sum is the total misclassification percentage
> 
> first/second set of brackets seem to be evaluation statistics of each 
> rule on (non-current) [current] class

The second set of brackets are stats computed on the hold-out set that 
is used for pruning the tree.

Cheers,
Mark.

_______________________________________________
Wekalist mailing list
Send posts to: Wekalist <at> 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

238	May 2013
358	April 2013
305	March 2013
302	February 2013
216	January 2013
200	December 2012
174	November 2012
178	October 2012
121	September 2012
180	August 2012
147	July 2012
192	June 2012
228	May 2012
220	April 2012
335	March 2012
233	February 2012
224	January 2012
285	December 2011
284	November 2011
247	October 2011
247	September 2011
326	August 2011
275	July 2011
249	June 2011
320	May 2011
274	April 2011
461	March 2011
365	February 2011
300	January 2011
234	December 2010
229	November 2010
338	October 2010
289	September 2010
349	August 2010
271	July 2010
345	June 2010
410	May 2010
336	April 2010
302	March 2010
264	February 2010
397	January 2010
332	December 2009
365	November 2009
362	October 2009
436	September 2009
453	August 2009
317	July 2009
455	June 2009
527	May 2009
501	April 2009
525	March 2009
362	February 2009
336	January 2009
222	December 2008
338	November 2008
331	October 2008
216	September 2008
250	August 2008
325	July 2008
282	June 2008
219	May 2008
354	April 2008
338	March 2008
264	February 2008
262	January 2008
197	December 2007
214	November 2007
288	October 2007
245	September 2007
176	August 2007
311	July 2007
276	June 2007
363	May 2007
410	April 2007
302	March 2007
231	February 2007
236	January 2007
155	December 2006
262	November 2006
197	October 2006
199	September 2006
188	August 2006
207	July 2006
289	June 2006
315	May 2006
289	April 2006
330	March 2006
238	February 2006
198	January 2006
212	December 2005
206	November 2005
253	October 2005
280	September 2005
271	August 2005
157	July 2005
267	June 2005
221	May 2005
286	April 2005
182	March 2005
117	February 2005
111	January 2005
108	December 2004
106	November 2004
81	October 2004
56	September 2004
69	August 2004
101	July 2004
143	June 2004
96	May 2004
87	April 2004
71	March 2004
104	February 2004
96	January 2004
70	December 2003
143	November 2003
97	October 2003
80	September 2003
87	August 2003
99	July 2003
90	June 2003
100	May 2003
123	April 2003
122	March 2003
119	February 2003
94	January 2003
40	December 2002
137	November 2002
122	October 2002
81	September 2002
55	August 2002
61	July 2002
43	June 2002
70	May 2002
35	April 2002
3	January 2001
2	March 2000

Mon	Tue	Wed	Thu	Fri	Sat	Sun

				1	2	3
4	5	6	7	8	9	10
11	12	13	14	15	16	17
18	19	20	21	22	23	24
25	26	27	28	29	30	31