05 Support Vector Machines Flashcards
What are SMV’s used for?
Classification, Regression, Outlier Detection
What size of dataset are SVM best suited for?
it is small & medium data sets
Can SVM be used for both linear and non-linear problems
Yes
How does SVM classify?
In any classification we need the model to separate classes but we also want the model to have a distance from decision boundary as new instance can be added in the boundary which may lead to misclassification and SVM will maintain margin from decision boundary as well.
Do SVM need scaling?
Yes, we need to use standard scaling.
What is hard margin classification?
when we are strict about boundaries being wide and instances being on the right side of the boundary.
what is soft margin classification?
the street is kept as large as possible and margin violations are limited.
which hyperparameter is responsible to control the margin violation?
C - if the value is less the width of the street is wide maximizing margin violation and more value of c indicates that the width of street is less hence minimizing the margin violation.
how to deal with overfitting in SVM model?
we need to reduce the value of C.
Does SVM model give probability?
No SVM does not give probability.
how can we solve polynomial equation using SVM?
- we can add a new feature according to the degree.
- if there are m instances, we can add m features to the dataset.
what to do when we have a small dataset but the polynomial equation with high degree?
we can use SVM with kernel hyperparameter set to - “poly”
how can we add m features to the datasets if there are m instances in the dataset?
this can be done in svm using rbf (Radial Basis Function) kernel.
what are the hyperparameters f RBF kernel?
it has gamma and c parameter. both are used for the smoothness of the decision boundary.
when there is overfitting - we need to decrease the value of gamma and when there is underfitting we need to increase the value of gamma.
how is svc different from svr?
in svr we the width of the margin does not impact the prediction of the model but violation of the margin should not happen.