One-vs-one with SVM in Classification Learner

Hello,

I have tried to use Classification Learner with Quadratic kernel SVM to classify the data that include four classes, Normal, Low, Medium and High. So, If I choose one-vs-one for binary class, I will receive 6 binary classifiers from (k*(k-1))/2.

First question, the pattern of 1 binary classifier will be similar in the equation in the picture, right?. So, 6 classifiers is High-Low, High-Medium, High-Normal, Low-Medium,Low-Normal, and Medium-Normal ,right?

Second question, I would like to know that how one-vs-one of Classification Learner work to predict class from 6 binary classifiers.

Thank you

Answer 1

[Anay]

Hi Nuchaba,

Your approach is correct. If you choose one-versus-one Coding Design, you will have “K(K – 1)/2” binary learners for “K” number of classes.

“How Classification Learner predicts class?” depends on the decoding scheme. You can use either the “Loss-based” decoding which means that the predicted class will be the one which has the lowest average binary loss across all the binary learners or the “Loss-weighted” decoding which means that the predicted class will be the one with lowest average binary loss across the binary learners corresponding to that class.

You can find more information on the Coding Design, binary loss in the MATLAB documentation by following the below steps:

1. In your MTLAB command window, enter the following commands:

>>doc fitcecoc

This will open documentation for the “fitcecoc” function

2. Navigate to the More About section to find documentation of the Binary Loss and Coding Design.

I hope this helps!