Precision - Recall: Graphical Representation

Question

I'm a little bit confused with precision recall. I read some papers about recommender systems, where in one paper they have a graphical representation and in other papers they don't (they just have the values for precision/recall). What is the advantage of a graphical representation?

Am I right, that for a graph you have to do something like a step function? e.g. you have a list of values true,true,false,true,false and you begin with the first one, calculate precision recall and then draw that point into the graph. For the next point, you take the first and the second value true,true into account, calculate precision recall, draw the point... and so on. Am I right that this is similar to drawing a ROC curve?

Answer 1

The process of creating a precision-recall curve is very similar to creating a ROC curve, i.e. in both cases your predictions have to be associated with some kind of score such as a probability for each class. As for ROC curves you gradually increase the threshold on your score and each time calculate two performance measures based on the resulting confusion matrix when applying the current threshold on the score. For ROC curves you calculate sensitivity (y-axis) and 1 - specificity (x-axis), whereas for precision-recall curve you calculate precision/sensitivity (x-axis) and recall (y-axis). As you can see both ROC curves and precision-recall curves use the sensitivity, but compare it with different value.

If the paper you were talking about only mentions a single value for precision and recall it might be that their classifier does not return continuous scores, but only the predicted class, hence you cannot construct a precision-recall curve.