machine learning - Choosing an operating point with SVM's, and categorical models in general -
so, i'm trying classify data belonging 1 of 2 classes. classes mutually exclusive, , every data point belongs 1 or other.
so tried few classifiers. in case of svm, 2 outputs, probability of belonging category a, , probability of belonging category b. in type of model, believe normal course of action winner takes all.
that doesn't give me flexibility in terms of choosing operating point, based on sensitivity/specificity requirements. in mind, opted use p(b) - p(a) score.
intuitively, difference makes sense me. if model 100% sure data point category b, , 100% sure it's not category a, score of 1. in reverse situation, score of -1.
since did make on whim, however, i'm not sure best way turn categorical model binary decision. seem work, , outperform neural networks i've been training up. thought i'd ask. thoughts?
first of operating point mean setting threshold? .
even though getting better fit doesn't mean good. p(b)-p(a) seems fine in case mentioned. when p(a) =0.7 , p(3)=0.4 ? . clear point belongs class a. subtracting classified in wrong class. suggestion use direct classification results.
Comments
Post a Comment