Scribd
Upload
32 views

Determining Classification Error: FAR Vs 1-FRR

This document discusses determining classification error rates by calculating false acceptance rate (FAR) and false rejection rate (FRR) based on a contingency table. It provides an example where the algorithm classifies samples s2 and s3 as outliers when they are not, resulting in false positives. Sample s9 is not detected as an outlier when it is, resulting in a false negative. Based on this, the contingency table is constructed and FAR and FRR are calculated for this run. Pseudocode is provided to loop through multiple parameter values, calculate FAR and 1-FRR each time, and output the results to a file to later plot.

Uploaded by

Anas Melhem
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
32 views

Determining Classification Error: FAR Vs 1-FRR

This document discusses determining classification error rates by calculating false acceptance rate (FAR) and false rejection rate (FRR) based on a contingency table. It provides an example where the algorithm classifies samples s2 and s3 as outliers when they are not, resulting in false positives. Sample s9 is not detected as an outlier when it is, resulting in a false negative. Based on this, the contingency table is constructed and FAR and FRR are calculated for this run. Pseudocode is provided to loop through multiple parameter values, calculate FAR and 1-FRR each time, and output the results to a file to later plot.

Uploaded by

Anas Melhem
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 2

Determining classification error: FAR vs 1-FRR

Problem: Given samples s1, s2, , s10, we try to detect outliers. How good is our
algorithm?
Given specific values for d and p, our program classifies some samples as outliers, and
some as not outliers. (In other words, the number of output classes is 2; so the
contingency table will have dimension 2x2).
The true outliers are s8 and s9.
Find the error.
The contingency table will look like this:
True results
outliers
Our
programs
results

not outliers

outliers
not outliers

10
Number of false positives = FP (is the number in first row, column 2)
Number of false negatives = FN (the number in second row, column 1)
Total number = TN = 10, always (because we have 10 samples).
FAR = FP/TN
FRR = FN/TN
Every time you run your program for a certain value of p and a certain value of d, there
will be a new contingency table.
For example, lets assume we ran the program for d=4 and p =3, and our programs
output is that the outliers are s2, s3, and s8. So, our program detected s2 and s3 as outliers
but they are not outliers; therefore, s2 and s3 are false positives. Our program did not
detect s9 as outlier although it is an outlier, so s9 is a false negative. For this run, the
contingency table is:

Our
programs
results

outliers

True results
outliers
1 (i.e. s8)

Not outliers

1 (i.e. s9)
2

FP = 2
FN = 1

Not outliers
2 (i.e. s2 and 3
s3)
6 (i.e. s1, s4,
7
s5, s6, s7, and
s10)
8
10

TN = 10
FAR = 2/10
FRR = 1/10
Output FAR vs 1-FRR into a file, so that it can be plotted later.
Pseudocode:
//Assume that your program stores its results (i.e. the detected outliers) into an array
output[].
//Assume that the true outliers are in array trueouts[], where trueouts[1]=s8 and
trueouts[2] = s9.
//Assume TN = 10
ErrorPlot() {
Open file toBploted for appended writing
For (several (p,d) values) {
FP=0
For i=1, sizeof(output[])
If output[i]!=s8 or s9,
FP++
}
}
FN=0
For (i=1, sizeof(trueouts[]) {
If trueouts[i] is not found in output[] {
FN++
}
}
FAR = FP/TN
FRR = FN/TN
Print FAR, 1-FRR into toBploted
}
}

You might also like