Comparison of Support Vector Results with Non linear regression results - Statistica General Discussion - Statistica - Dell Community

Comparison of Support Vector Results with Non linear regression results

Comparison of Support Vector Results with Non linear regression results

This question is not answered

Hi,

I used SVM functionality of Statistica in a research. One of the reviewers requested comparison of SVM results with Non-Linear Regression results using these words "A nonlinear solution will almost always be at least as good as a linear solution. However, you need to justify SVM in contrast to other nonlinear regression approaches."  

From this statement I understood that I should execute a similar test with the same variables using non-linear regression menu. I checked the menu Statistics/Advanced Models /Fixed Non-Linear Regression Seemed proper to me. ( I am not an expert on statistics as you would understand). However this menu is designed in a completely different manner than SVM. For example, I can not select test and training data, or I can not use n-fold cross validation. 

So how should I compare results of these two analyses? For SVM I included graphs with fitting lines and it was found ok by the reviewers for non-linear regression what output should I include to my research text? 

You may suggest different steps to fulfill the reviewer's request?  

Thank you for your time and consideration. 

Ferda 

 

. A nonlinear solution will almost always be at least as good as a linear solution. However, you need to justify SVM in contrast to other nonlinear regression approaches.

All Replies
  • Hi Ferda,

    The general purpose for Fixed Nonlinear Regression module is to specify nonlinear transformations, and then use these transformed variables in your regression analysis.Here links Fixed Nonlinear Regression - Example and our help document about Fixed Nonlinear Regression (also available in the e-manual). There is currently no n-fold cross validation feature embedded in this particular module. However, you can use Data|Sampling|Split node random sampling to split your data into testing and training data. 

    For predicting new observations, there is Residuals/assumptions/prediction tab under the results dialog where you can specify values for independent variables to predict dependent variable. For predicting cases in a testing set, you can apply the estimated coefficients obtained from the regression result to a computation formula on the variables in a spreadsheet. Here is how to apply a formula in a spreadsheet.

    Unfortunately, we are not legally defended to offer advises regarding how you should compare your results. We are here to assist you with software related questions but can not provide statistical consulting at standard support service. 

  • Hi Ferda,

    I'm not sure if this answers your question or not but we did have a Tech Webinar on Evaluating Model Quality if it helps.

    software.dell.com/.../tech-webcast-evaluating-model-quality891560