Evaluation of Prediction Accuracy

Let us suggest that for compounds in ES we have values of some targeted molecular property. Expert divides ES into two parts positive and negative examples. Using a constructed estimator we calculate P(C) values and, selecting the threshold value, divide ES into two other parts predicted positive 

For pattern recognition or classification, usually, the following characteristics of recognition accuracy are used (see, for example, refs. 66,106-108)  

Accuracy (concordance) Predictive value positive Predictive value negative False negative rate  

Estimation of the optimal threshold value is provided by minimizing a risk function, which depends on a priori probabilities of positive and negative 

In any case, this approach uses several additional assumptions. For this reason in the last time in ML the recognition accuracy criterion of the Area Under the ROC Curve AUC), which is free of additional assumptions, becomes very popular. Mathematically, AUC equals the 

---