Kernel Machines Blackboard - All Forums

help on data format: matrices

schokoholic — Sat, 21 Nov 2009 17:26:22 +0000

Hi everyone,

i should use libsvm for my thesis work, which consist in using machine learning to predict the outcome of a maximux clique problem.

now, there is my problem:

each instance of the training data file is a matrices with associate a label (+1/-1)

i've read that the data format is: : ... etc...

is right the to construe "value" as the numbers which compose the matrice, and "index" as the position of the namber in the matrice???

or better which is the proper way to rapresent a matrice, according with the data format??

hope that i explained myself clearly and that somebody can help me doing this.

thank you all in advance.

schoko

libsvm and liblinier crash

BezeLyE — Sat, 21 Nov 2009 14:37:38 +0000

hi,

i'm trying to use netflixprize dataset.. may be you don't know it has 17770 movies rated by ~480k user and about ~1.7m rates..

my raw problem file likes below

1: 2:
5 1:34452 2:859
3 1:22453 2:11230
4 1:99223 2:122
....
and i scaled this file with svm-scale.exe
and it has ~1.7m rows
and it is ~1.8gb (huge i guess)

i have tried both libsvm(svm-train.exe) and liblinear (train.exe) with default parameters. every time i try they both crashed..
i have 2gb memory i thought there may be a memory crash
so i decided to use half of the data.. again both libsvm and liblinear crashed..

is there any limitations like vector count?
does anybody know why they are crashing?
are libsvm and liblinear suitable for this problem?
i am trying to guess the rate which user didn't rate that movie but has rates given other movies..
thanks..

LIBSVM, matlab x64 running issues

ungh — Fri, 20 Nov 2009 23:52:24 +0000

Hi.

Having an issue compiling/running libsvm-mat v2.9 under Windows

OS: Windows XP, 64 bit
Matlab 7.7.0 (R2008b)
Compiler:
[1] Microsoft Visual C++ 2008 Express in C:\Program Files (x86)\Microsoft Visual Studio 9.0

I did mex -setup to set the compiler.
make.m gives
"'mt' is not recognized as an internal or external command,
operable program or batch file."

for lines
mex -O svmtrain.c svm.obj svm_model_matlab.obj
mex -O svmpredict.c svm.obj svm_model_matlab.obj
mex -O libsvmread.c
mex -O libsvmwrite.c

but it still produces the .obj and .mexw64

But svmtrain -> "??? Undefined function or variable 'svmtrain'."

I've added -largeArrayDims and checked that I have the 64bit compiler cl.exe in amd64 within visual studio express.

Any ideas?

Thanks
ungh

Help: liblinear takes forever!

logicators — Thu, 19 Nov 2009 00:35:59 +0000

Hi,

I found that liblinear may take an unpredictable amount of time depending on the number of features, classes, and instances. For example, on a text classification problem with about 55K features and about 40K instances (and about 200 classes), it ran for about 8 hours and did not converge. I tried switching to the primal version and also tried the multiclass option without any luck. I also tried increasing the eps to 1.0 and it still ran for hours.

This problem is not unique to this specific dataset. On many standard text datasets (such as the ones that ship with the Cluto clustering toolkit), the amount of time it takes is highly unpredictable, and it doesn't seem to be running in linear time.

The SVM guide says that if number of features is substantially less than the number of instances, switching to primal should resolve the performance problem. But I found that there is no ratio between the number of instances and features that can be used to reliably switch between dual and primal versions. If I switch to primal version on a dataset with twice as many instances as features, it may reduce the runtime on that dataset but not on another dataset.

Any help is highly appreciated!

Who can help me with the data-set selection?

wangyoang — Wed, 18 Nov 2009 02:37:53 +0000

Now I have some data, with the format: 20-dimensional vector, and a score(between 0~1). Now I want to get a svm-model which can identify the data which score is higher than 0.9 as correctly as possible. How should I select the training-set?

many thanks.

Any existing tool for multi-label classification?

need47 — Mon, 16 Nov 2009 22:33:26 +0000

Hi, folks

This post is a continuation of my previous one:
http://agbs.kyb.tuebingen.mpg.de/km/bb/s...p?tid=1374

I realize now my problem is actually multi-label classification, not multi-class classification. I googled and found some algorithms. But I'm new and not familiar with those theories, so it's hard for me to re-implement by coding.

Please anyone suggest a existing tool? Thanks in advance.

Jay

basic question please urgent

bigfoot — Fri, 13 Nov 2009 21:38:51 +0000

Hi all

I have a problem testing new equation satisfy mercer's condition which state that all values in the matrix are >0 that means the created K matrix has to be semi-positive definite, i.e. it has non-negative
eigenvalues.

but when i tested the well known function such as polynomial function it gave me a negative eigenvalue the simple code i used to check whether right or rong is

>> dd=eig(K); % K = (X*X' + 1)^2; since X is the input matrix to be verified
>> ispos=dd>0;

the output looks like

> 1
> 1
> 1
> 1
> 1
> 1
> 1
> 1
> 1
> 1
> 1
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0
> 0

please any help regarding that i do appreciate it..

how do I run LibSVM??

schokoholic — Fri, 13 Nov 2009 21:29:52 +0000

hi guys,
i might be stupid, but i can't understand how to launch libsvm software without using matlab version....

I'm using win7!

can someone help me please... it is very important! i need it to work for my thesis!

thank you all

multiple dependents with SVM

need47 — Fri, 13 Nov 2009 14:34:11 +0000

Hi folks,

As I know, SVM toolkit such as libsvm can successfully deal with problem of only one dependent (The typical classification/regression belongs to this category). However, my problem has k dependents:

y11, y12, ..., y1k -> x11, x12, ..., x1n
y21, y22, ..., y2k -> x21, x22, ..., x2n
.... -> ....

How can I build a model using SVM method? Or is there any better solution that can handle it?

Many thanks.

How to convert other data formats to LIBSVM format?

Amherstclane — Fri, 13 Nov 2009 09:58:29 +0000

It seems that I have become addicted to finding my answers on the web.
How to convert other data formats to LIBSVM format?
I am completely new to this.
Any suggestions you give me will be helpful.
Thanks.

R2W2 for multi-class in Spider

klonona — Mon, 09 Nov 2009 10:27:35 +0000

Hello all,

I am new to SVM and I found this forum very helpful. I am using the SPIDER toolbox currently and after reading the doc from the SPIDER website and the posts here, I am able to conduct the feature selection with RFE for multiclass problem (well, one vs rest and mc_svm work, one vs one has some problem). I tried to use other feature selection method, r2w2_sel, and found that it seems doesnt support the multi-class problem. I wonder if anyone tried to make it work or is there any simple modification I could try to make it work?

Another question, if I have
a = rfe
a.feat = 10
a.output_rank = 1
...
[r a] = train(a, data)
what does a.rank really gives me?
Say, I have 30 features in total, I see that a.rank seems to rank me the importance of the features, but what does it do to the feature number (in my case, 10)?

Thanks a lot, I appreciate your help.
what i

Linking Error How to solve??

stuartjems — Mon, 09 Nov 2009 07:36:51 +0000

Hello Everyone,

I need help.I have some linking error in my c++ program.
I am working on c++ on Linux.I have included all need header files required.
But there is a problem of linking error.
How to solve it?

Can anyone help me out??How to solve this linking error?

libsvm python how to do regression?

iamthegoodbot — Sun, 08 Nov 2009 23:49:52 +0000

script told me:

Code:

Model doesn't contain information for SVR probability inference

for

Code:

        param = svm_parameter(kernel_type = RBF, C = 10)

        prob = svm_problem(labels, points)

        m2 = svm_model(prob, param)

what's wrong? [/code]

Question about cross validation in libsvm

Eman — Sun, 08 Nov 2009 15:50:51 +0000

Hello,

I would like to know what cross validation is used in libsvm?
Is it k-fold or one leave out or hold-out ..?
and are the samples chosen randomly or in sequence?

Thanks

how to use grid.py to get the optimal parameter values?

Eman — Sun, 08 Nov 2009 15:16:26 +0000

Hello,

I don't know anything about python, and I want to use grid.py and easy.py to get the best paramters and apply scaling.
What are the tools I need to install ? and how can I run it?

Thanks alot

Libsvm: c & g values and overtraining

Dale — Sun, 08 Nov 2009 14:43:40 +0000

Short question,
when I get my best c-values = ~32000 and g = ~0,0001 is this a hint of overtraining or completely normal? I ask because my cross-validation results and unseen test data results can sometimes be very different in accuracy.

My data is balanced and has around 3000 observations with 70 features and 3 classes.

Thanks!

libsvm: cross validation result VS. unseen test data result

Dale — Sun, 08 Nov 2009 13:46:19 +0000

Hello!

So here are my results from running easy.py:
Cross-validation (3000 observations, multiclass): 78 % accuracy
Test data / unseen data (600 observations, multiclass): 45 % accuracy

Could somebody explain to me why I get so very different results from these tests? Both datasets originate from the same dataset (they've been split up randomly with subset.py) and contain around 70 features. The data is balanced.

LibSVM: Decision values??

Eman — Sun, 08 Nov 2009 03:14:54 +0000

Hi,

I am using LibSVM matlab version 2.89 , I was training the model on binary classification problem when I found that decision values always result in negative value ..

I would like to ask : what is the relation between predict_label and dec_values?? Isn't it that predict_label = sign(dec_values)??

Why do I always get negative value?

Thanks alot

About Naive Bayes

mcad — Sat, 07 Nov 2009 17:24:03 +0000

Hi,
I heard that Naive Bayes method can be much mor eefficient with numerical
attributes rather than categorical attributes. However, I could not
understand the reason for this. Would anyone explain this to me please ?

Thanks in advance

Convert CSV to LIBSVM

Dale — Sat, 07 Nov 2009 11:44:19 +0000

How do I make this conversion? I've actually done it successfully in the past using WEKA, however this current data set I'm unable to make the conversion. The option to "save as LIBSVM" just doesn't show up in the SAVE menu as it's done in the past.

I found this thread: http://old.nabble.com/converting-CSV-dat...02227.html but even when all my attributes are numeric and I set one of the columns as a class... it still doesn't work.

So, could somebody please walk me through the process of making the conversion? Can it be done in WEKA? Matlab? Python? Java? Please provide code if you have it readily available!

Thank you!

Update. Got it working with, ehum, reading the FAQ...

I.e.,
matlab> SPECTF = csvread('SPECTF.train'); % read a csv file
matlab> labels = SPECTF(:, 1); % labels from the 1st column
matlab> features = SPECTF(:, 2:end);
matlab> features_sparse = sparse(features); % features must be in a sparse matrix
matlab> libsvmwrite('SPECTFlibsvm.train', labels, features_sparse);

However, to my defense... I've had much trouble setting up the compiler in Matlab but this time it just (magically) worked

Still, why didn't it work in WEKA?