The performance metrics used are: accuracy rate , recall , F. . 1 , TPR , the FPR .
| Prediction is true | Prediction is false | |
|---|---|---|
| True true | TP(true positive) | FN(false negative) |
| True false | FP(false positive) | TN(true negative) |
Precision ratio Precision = TP / (TP + FP )
Recall Recall = TP / (TP + FN )
Example real rate is the probability of positive examples is determined as positive examples of TPR = TP / (TP + FN)
Example false positive rate is the probability of the counterexample is determined to be positive examples FPR = FP / (TN + FP )
Precision accuracy rate, also known as the name suggests is suitable for high accuracy applications such as web crawlers and recommendation. Recall also known as the recall ratio for the detection of credit risk, fugitive information. Precision and recall are a contradictory measure, it is necessary to find a balance point , tend to use F. . 1 is a harmonic mean of precision and recall:
(1) error rate and accuracy
Error rate:
Accuracy: acc = 1-e
(2) the AUC and ROC curves
For classification 0,1, some of the results obtained classifier is not 0 or 1, such as neural networks and the like obtained is 0.5, 0.6, then you need a threshold of the cutoff , then the normalized value is less than the threshold 0, is greater than a normalized 1, you can get a classification result.
ROC curve (Receiver Operational Characteristic Curve) False Positive Rate is the abscissa, True Postive Rate curve plotted as ordinate.
Point in the graph represents the sensitivity and specificity balance between, for example the left, that is, the smaller the false positive, true positive is also smaller. The larger the area under the curve, it indicates that the method is more advantageous to distinguish two categories.
AUC is the ROC curve covered area of .
AI interview questions more public attention number: We are all code Agriculture (allmanong), or scan the next Fanger Wei code!
