As per the documentation of 'fitcsvm', the parameter 'ClipAlphas' is available and set to true as a defualt. It mentions if it is set to true then, at each iteration, if αj is near 0 or near Cj, then MATLAB sets αj to 0 or to Cj, respectively.
Potential reasons for clipping alpha values:
- This would help in numerical stability during the optimization process as in some cases the optimization algorithm might produce very large or very small alpha values leading to instability.
- As correctly identified, it helps the algorithm to achieve a faster convergence by limiting the the range of alpha values, prevents it from reaching extreme values and prevents overfitting.
- In Case of SVM, alpha clipping would ensure that no single support vector has an excessively large influence on the decision boundary.
Incase, your use-case would require fine-tuning the SVM model without clipping alpha values i.e. you have a non-standard optimization requirement which could require controlling alpha value carefully where numerical stability will not be an issue, the setting the 'ClipAlphas' property to false can be an option to exercise.
The following Mathworks documentation can be referred to know more:
'ClipAlphas' (fitcsvm): https://www.mathworks.com/help/stats/fitcsvm.html#bt9w6j6_sep_shared-ClipAlphas
Thanks.
