Fix `predict.mildsvm()` cutoffs
skent259 opened this issue · comments
In practice, the predict method scores for mildsvm may be far off from the cutoff of 0. This means that it may be better to come up with a data-driven cutoff or add parameters to let the user specify the cutoff. Otherwise, there is a risk of poor performance
I think this issue largely stems from a poor choice of weights. mildsvm()
has improved default weights now, which should alleviate part of this problem. Look into further for 0.3.0
Youden's J Statistic may provide for an optimal cutoff: https://en.wikipedia.org/wiki/Youden%27s_J_statistic
After further investigation, I'm not sure that this is a major issue. Users always have the option to pull the scores directly and adjust their cutoff manually. Any choice of better cutoff would be ad-hoc in this code