Is SVM sensitive to unbalanced observations? The observations in one class is 3-4 times of the observation in an other class in binary classification

My first question is sensitivity of SVM to unbalanced datapoints. How much SVM is sensitive to that?

And is there any functionality designed i the fitcsvm to account for the unbalance in the datapoints in binary classification? I know that oversampling the smaller class or undersampling the larger class can be a solution to deal with "unbalanced" observation but I am interested for other approaches.

I checke "prior" and found it's role is only to remove observations with zero prior probablity and apparently doesnot play role in the classification step.

Answer 1

[Prince Kumar]

Hi Zeynab Mousavikhamene,

Yes, SVM is sensitive to imbalanced dataset and this gives suboptimal models.

You can use 'Cost' Name-Value pair and pass a cost matrix. fitcsvm uses the input cost matrix to adjust the prior class probabilities.

You can refer the following link for more information :

https://www.mathworks.com/help/stats/fitcsvm.html#namevaluepairs