My problem is multiclassification. It has four classes.
I know that using one-one algorithm 6 binary classes will be formed and for one to rest algorithm 4 binjary classes will be there.
Libsvm is using 0ne-one approach.
When I use LIBSVM for classifiaction task it is showing last few lines as follows:
optimization finished, #iter = 553
nu = 0.100522
obj = -1480.765685, rho = -1.798735
nSV = 58, nBSV = 19
Total nSV = 58
Accuracy = 96.4539% (136/141) (classification)
Is this output and accuracy is the final overall output of the model? or it is the output of last binary class?
Also, here accuracy means F-score?
Dear Krupal,
obj is the optimal objective value of the dual SVM problem. rho is the bias term in the decision function sgn(w^Tx - rho). nSV and nBSV are number of support vectors and bounded support vectors (i.e., alpha_i = C). nu-svm is a somewhat equivalent form of C-SVM where C is replaced by nu. nu simply shows the corresponding parameter. Accuracy is # correctly predicted data / # total testing data × 100%. More details are in the following document
http://www.csie.ntu.edu.tw/~cjlin/papers/libsvm.pdf
Thanks for your reply Gino.
But, I think you don't understand my question or may be I have not explained it properely. Actually I know interpretation of each parameter which you have given.
My probelm is I want to find value of decision function in my work. So, I need to know support vectors. I want to clarify that , when I give model.SVs, whatever support vectors it is displaying is it for my final output? or I have to incorporate last 4 values of support vectors and predictions in my work,as there are 4 binary classes(using one-rest approach) for 4 class problem?
Hi,
if your correctly predicted data 6 classes then,
Yes, this is output and accuracy is the final overall output of the model ,
@Khamael
Thanks.
Are you sure the last output is the overall output for my classifier.?Not for only last choosen class.
It means the last few lines it is displaying the final output .so, I can proceede taking the last output.
Hello Krupal
Actually my problem is also related to multiclass classification.
But I have one doubt. When we are making 6 binary classes then for each binary class we have to do prediction i.e,the svm-train and svm-predict commands, we have to run 6 times (i.e one time for one classifier) then how will we get to know the prediction accuracy.
Please explain .
Thanks a lot for your help.
Hi Jaya,
If you are using LIBSVM,then during training set parameter value -s 0, which will take care of multiclassification. So, it will create k*(k-1)/2 classes and train each class and give the output after training each class.So, you don't have to run 6 times. Run it only once.The predict accuracy in the final output is overall output considering all 6 classes. When you supply data file, it should have data of all six classes
Hi,
I am also working on libsvm but i am kind of new in SVMs.
What i am trying to find out is to analyse normal and abnormal vibrations. So I have 2 sets of data, one with normal condition and the other one is the abnormal condition. What I would like to do is to train both of these conditions using SVM algorithm, and when i introduce a new set of data, it should understand if it is normal or abnormal. Can you please assist me on this including how exactly should i input my data?
My current data format is as follows (is it ok for this application or is there a need for more than 2 columns? Kind of confused there):
Normal Condition:
...
+1 1:0.112
+1 1:0.049
+1 1:-0.027
+1 1:-0.110
...
Abnormal Condition:
...
-1 1:0.295
-1 1:0.120
-1 1:0.017
-1 1:0.215
-1 1:0.281
-1 1:0.178
-1 1:-0.095
...
Krupal,
The “one against all” strategy consists of building one SVM per class, trained to identify the instances in a single class from the instances in all remaining classes. This is the most popular approach for SVM.
In your task, seems like you're using the "nu" parameter of SVM, so both output and accuracy (you had) are the final overall output of the model that gets evaluated using the selected evaluation method. Meaning that, if you have x number of classes, this model is going to predict them all.
However, for your second question, accuracy means the percentage of correctly classified instances (i.e., accuracy of your classifier model), while F-score is known as F-measure.
HTH.
Samer
Fatih,
If you're going to predict any of the values you highlighted up, then you're working with a regression problem. Thus, you need to use another type of SVM that handles numeric classes, and this requires selecting suitable parameter (i.e., either epsilon or nu) .
HTH.
Samer
Hello everyone,
I'm working on a project which stores accelerometer values of a user in a database for a certain period of time (roughly 5-6 hours). I'm working on Android Studio LibSVM. I need to classify a user as normal/abnormal based on Z values of the accelerometer. How should my data input be? Further, with the accuracy and other terms (in the output format), how do I actually come to know the classification as normal/abnormal?
Any help would be appreciated
Thank you
Hi Siddhi,
Please see below.
Q1: How should my data input be?
Answer:
Use the output of the Z values as an input of the algorithm.
Q2: how do I actually come to know the classification as normal/abnormal?
Answer:
Accuracy and other metrics give you an overall idea about the performance of your predictive model. (Not predicting one target label in particular. Such metrics let you know how well the algorithm predicts the class you have).
HTH.
Samer, PhD.
hello Krupal,
please suggest me how to use this output of the svmtrain for prediction using svmpredict?
ptimization finished, #iter = 3342
nu = 0.027871
obj = -709.872334, rho = -0.086839
nSV = 867, nBSV = 1
Total nSV = 867
Cross Validation Accuracy = 99.6461%
You just use svmpredict function for test samples. It will give to the accuracy result of your test sample.
Hi can you suggest how do I construct a hyperplane using the LibSVM with RBF kernel?