We use cosine or Euclidean distances for embedding models to make hard classifications. But this has a big limitation: no measure of confidence and no interpretability.
Instead, building a logistic regression model can turn distances into percentage based confidence scores. This also accounted for how a small decrease in distance affects the confidence score—similar to how a derivative measures sensitivity.
Instead, building a logistic regression model can turn distances into percentage based confidence scores. This also accounted for how a small decrease in distance affects the confidence score—similar to how a derivative measures sensitivity.