In sklearn.svm.SVC the default value of the parameter gamma is 'scale', i.e. gamma = 1 / (n_features * X.var()). What is the explanation for this default choice of gamma and why does it work so well (at least for my dataset, I couldn't beat this value with extensive grid-search for gamma)?
For parameter tuning, you shouldn’t concentrate only on obtaining the highest accuracy. Are you talking about the accuracy of your model only or you have validated it by other criteria as well?
https://stats.stackexchange.com/questions/499916/why-svm-with-gamma-scale-for-rbf-kernel-works-so-well#:~:text=The%20intuitive%20explanation%20for%20the,high%20values%20meaning%20'close'.
Normally in SVM using RBF kernel both gamma and the cost are tuned to create separate classes. The radial basis part of the name comes from the fact that this function decreases in value as it moves away from the center. The center in this case being the support-vector. The gamma parameter makes the most intuitive sense when we think about the RBF (or Gaussian) kernel. The Gaussian class boundaries dissipate as they get further from the support vectors. The gamma parameter determines how quickly this dissipation happens; larger values decrease the effect of any individual support-vector.
In brief sometimes the high you make gamma value may lead to the right classification (exact classes) or over classification. The lower gamma is may lead to under classification. Also, you may notice sometimes after the classification process and after increasing the value of Gamma parameters that the accuracy is high.
- Badges
- Science topic