1. four cases
Precision accuracy rate, Recall recall, is a common evaluation binary classification. Confusion matrix as follows:
Positive results were positive forecast | Negative predictions false positive | |
---|---|---|
True predictions to be true | TP | TN |
False predictions are false | FP | FN |
Often focus on class as positive class, other classes are negative category. (Dogs and cats to binary, for example, are concerned about the dog's precision and recall)
TP | Positive class prediction is positive category (predicted pictures of dogs dog is actually marked) |
---|---|
FN | Positive class prediction is negative category (cat pictures predicted actual label is a dog) |
FP | Negative class prediction is positive category (predicted pictures of dogs actually marked a cat) |
TN | Negative class prediction is negative category (cat pictures predicted label is actually a cat) |
T, F is the representative of the picture corresponding label
P, N is the representative of the picture predicted results
2. Precision
Accuracy rate is calculated:
\ [P = \ {FRAC the FP + TP TP} {} \]
understood that:
TP + FP: That is all Positive, which is predicted picture is the picture of the number of positive class
TP: That is also positive class is predicted to be positive picture of the number of classes
In short: the proportion of correct prediction of the number of pictures of the total number of positive class prediction (prediction from the perspective of how many predictions are accurate)
3. Recall
Recall formula:
\ [R & lt = \ {FRAC TP TP + FN} {} \]
understood that:
TP + FN: that is, to fully meet all of the picture marked number of pictures
TP: positive class is predicted to be the number of positive class picture
In short: Determine the positive class is predicted to account for the number of all positive class pictures marked images (from a label perspective, how many were recalled)
Example 4. dichotomous
Or in cats and dogs binary, for example, the test set a total of 20 dogs, 20 cats pictures marked picture (a dog as a positive example), the model predicts which has 16 dog pictures, 14 pictures marked indeed dog, leaving two pictures labeled as a cat.
Positive | Negative | All | |
---|---|---|---|
True | TP: 14 | TN: 6 | 20 |
False | FP: 2 | FN: | |
All | 16 |
It is possible to calculate a
\ [precision = \ frac {TP } {TP + FP} = \ frac {14} {14 + 2} \]
\[ recall = \frac{TP}{TP+FN} = \frac{14}{40} \]
Example 5. Multiple Classifiers
The reference from Example: https://www.itcodemonkey.com/article/9521.html
Class1 | Actual_Class1 | Actual_Class2 | Actual_Class3 |
---|---|---|---|
Predicted_Class1 | 30 | 20 | 10 |
Predicted_Class2 | 50 | 60 | 10 |
Predicted_Class3 | 20 | 20 | 80 |
For example, we calculate for class2:
class2-TP: label class2, predicted class2 = 60
class2-TN: label class2, prediction not class2 = 20 + 20 = 40
class2-FP: label is not class2, predicted class2 = 50 + 10 = 60
class2-FN: tag is not class2, predictions nor class2 = 30 + 10 + 20 + 80 = 140
6. Other indicators
F1 recall and precision values of the harmonic mean is:
\ [\ FRAC of F_1} {2} = {\ FRAC. 1} {P} + {\ R & lt FRAC {} {}. 1 \]
\[ F_1 = \frac{2TP}{2TP+FP+FN} \]