Top Page | Upper Page | Contents | About This Site | JAPANESE

Confusion Matrix

In Pattern Recognition , confusion matrix is used as the table of the output.
confusion matrix

Patterns of Confusion Matrix

For Perfect Judgment

If the system is perfect, there values only in upper left and lower right space.
output

For Perfect Mistake

If the values only in lower left and upper right space, the system mistakes perfectly.
output

Need More Data

If "only in upper left" or "only in lower right", they are not perfect system.

In the real world, there are not enough data for 4 parts of the confusion matrix.
output output

Evalation with Confusion Matrix

Differet fields use different ways for the evaluation with confusion matrix.

If Risk is Important

If risk is important, 2 types of risk is calcurated.
FP ratio = FP / (TN + FP)
FN ratio = FN / (TP + FN)

FP ratio and FN rario use same idea of Type I error and Type II error in Statistics .

If Positive is Importanr

FP ratio = FP / (TN + FP)
TP ratio = TP / (TP + FN)

This evaluation is used for screening. At screening, at first, TP ratio is maximized. Second, FP ratio is mimized.

ROC Curve and AUC are used for this evalation.

Relationship between Confusion Matrix and Threshold

If there is a data set and histgram below.
Output frequency histgram

For example, if threshold is "3", confusion matrix is below.
threshold is 3

For example, if threshold is "7", confusion matrix is below.
threshold is 7




NEXT ROC Curve and AUC