SVM cross validation parameters optimisation and accuracy

asked 2017-07-17 06:52:28 -0500

Nani gravatar image

I use the following code to train the svm using k-fold cross-validation but the prediction accuracy is low. What I am doing wrong and how to programmatically calculate the accuracy of the classifier using cross-validation.

Log.i(TAG,"Training..."); params.set_svm_type(CvSVM.C_SVC); params.set_kernel_type(CvSVM.RBF); params.set_C(1.0); params.set_degree(0.0); params.set_coef0(0.0); params.set_gamma(1.0); params.set_term_crit(new TermCriteria(TermCriteria.EPS, 10000, 1e-12));

    // k-fold cross validation
    int kFolds = 10;

    CvParamGrid C = new CvParamGrid();
    CvParamGrid p = new CvParamGrid();
    CvParamGrid nu = new CvParamGrid();
    CvParamGrid gamma = new CvParamGrid();
    CvParamGrid coeff = new CvParamGrid();
    CvParamGrid degree = new CvParamGrid();

    // initialize SVM object to avoid being Null object
    classifier = new CvSVM(trainingData, classes, new Mat(), new Mat(), params);

    classifier.train_auto(trainingData, classes, new Mat(), new Mat(), params, kFolds, C, gamma, p, nu, coeff, degree, false);;
    Log.i(TAG,"Training Done & Trained Model Saved");
edit retag flag offensive close merge delete


Well there is a plausible chance that you simply have not got enough data to train a better model? We need some more info on that first before we can make sure its due to the parameters.

StevenPuttemans gravatar imageStevenPuttemans ( 2017-07-18 08:06:37 -0500 )edit

Thanks @StevenPuttemans for your comment. I have 88 record with 40 negative images and 48 positive images. I suppose this due to the parameters as I have all the predictions of the test dataset 1.

Nani gravatar imageNani ( 2017-07-18 09:23:03 -0500 )edit

SVM on that small amount of data will never be optimal. Start by increasing your training set.

StevenPuttemans gravatar imageStevenPuttemans ( 2017-07-20 02:59:43 -0500 )edit

Thanks @StevenPuttemans, I will try to increase my data but my question is can I get at 90% accuracy with this data as we need it for real testing or it will be difficult?.

Nani gravatar imageNani ( 2017-07-20 03:12:42 -0500 )edit

Getting 90% certainty is always a challenge. But without more insight info in your application, making a random guess if it will work is impossible :D

StevenPuttemans gravatar imageStevenPuttemans ( 2017-07-20 03:17:37 -0500 )edit

The problem is that my colleague who work on matlab get nearly 97% accuracy with this data. Is the problem with the opencv and android as they are open source or other criteria. however, my app is a diagnosis app which give result of a disease based on analyzing colors in image.

Nani gravatar imageNani ( 2017-07-20 04:43:03 -0500 )edit

wait, if your collegue already trained with matlab, then you know the parameters right? Then you can simply set all these parameters and not add a grid for them?

StevenPuttemans gravatar imageStevenPuttemans ( 2017-07-20 07:12:09 -0500 )edit

Thanks, we will try this.

Nani gravatar imageNani ( 2017-07-20 07:15:35 -0500 )edit

Please suggest any solution I can try as we need to start testing from 1st of September and I have no enough time --> I am sorry, but this will not convince me further to provide support. Don't you have a professor or assistant that is assigned to this class?

StevenPuttemans gravatar imageStevenPuttemans ( 2017-07-24 06:30:29 -0500 )edit

Sorry for any inconvenience I caused. I will delete my comment as it seems inappropriate.

Nani gravatar imageNani ( 2017-07-24 08:20:29 -0500 )edit