PERFORMANCE OF A CLASSIFIER

DEFINITIONS

in a classification problem where class is a binary attribute the follow schema can be produced in order to study data

	POS-PRED	NEG-PRED
POS-TRUE	$TP$	$F{P}_{a-b}$
NEG-TRUE	$F{P}_{b-a}$	$TN$

where:

• $TP$ true positives
• $TN$ true negatives
• $FP$ false positives
• $FN$ false negatives

these measurements can be used to calculate some interesting performance metrics such as

• SUCCESS RATE (ACCURACY)

accuracy of the classifier

$$
\frac{TP + TN}{Ntest}
$$

• ERROR RATE

  $1-accuracy$

• PRECISION

  rate of true positives among positive classifications

  $$\frac{TP}{TP+FP}$$

• RECALL

  rate of positives that the classifier can catch (sensitivity)

  $$\frac{TP}{TP+FN}$$

• SPECIFICITY

  rate of negatives that the classifier can catch

  $$\frac{TN}{TN+FP}$$

• F1 SCORE

  armonic mean of precision and recall

  $$2\ast \frac{precision\ast recall}{precision+recall}$$

accuracy gives an inital idea of the performance but can be misleading when classes are unbalanced

f1 score is insteresting because is higher when precision and recall are balanced

if the cost of positive and negative errors are different than precision and recall should be considered

MULTI CLASS CASE

in a problem with non binary class attribute the previous table can be extended, it’s called confusion matrix

	a	b	c	Total
a	$T{P}_a$	$F{P}_{a-b}$	$F{P}_{c-a}$	$T_a$
b	$F{P}_{b-a}$	$T{P}_b$	$F{P}_{c-b}$	$T_b$
c	$F{P}_{c-a}$	$F{P}_{b-c}$	$T{P}_c$	$T_c$
Total	$P_a$	$P_b$	$P_c$	N

• $T_x$ true number of $x$ labels in the dataset

• $P_x$ total number of predictions of class $x$

• $T{P}_x$ true positives for class $x$

• $F{P}_{i-j}$ false positives for class $i$ predicted as $j$

• ACCURACY

  $$\frac{{\sum }_{i}TPi}{N}$$

• PRECISION $i$

  $$\frac{TPi}{Pi}$$

• RECALL $i$

  $$\frac{TPi}{Ti}$$

these measures can be global:

$$f(C)=\frac{\sum f(ci)}{C}$$

these measures can be weighted:

$$f(C)=\frac{\sum f(ci)\ast Ci}{C}$$

PREVIOUS NEXT