You are on page 1of 2

ROC curve

By:Ayush Shakya(069BCT505)

A Reciever Operating Curve(ROC) is a graphical plot of the performance of a classifier when its
discrimination threshold is varied.
If a classifier is to be represented with a histogram its plot may be as follows:

The discrimination threshold of 0.5 indicates that the test cases that the classifier assigns a score greater
than 0.5 is deemed to be a positive test case while the test cases with scores less than 0.5 are negative.
Sensitivity is the proportion of the test cases that are identified as positive when they are actually
positive(red area in the histogram or the true positive) while specificity is the proportion of the test
cases that are identified as negative while they're actually negative(blue area in the histogram or the
true negative).
From the histogram we can see that if we vary the discrimination threshold level(i.e move it to
the left or right), we see that the area covered by the sensitivity/specificity measure changes. More
precisely, if the discrimination threshold moves to the left, sensitivity increases and specificity
decreases. On the other hand, if the discrimination threshold moves to the left, sensitivity increases and
specificity decreases. Thus, sensitivity and specificity have an inverse relationship.
ROC curve is the plot of sensitivity vs (1specificity). The yellow dotted line is used
as a measure to determine the performance
of the classifier. If the ROC curve tends
towards the upper left corner of the graph it
is considered a good classifier whereas if
the curve ends of close the yellow line or
below its performance is considered bad.
Comparing the histogram and the ROC
curve, we can reach the following
conclusions:
a) for a near perfect classifier, such that the
two density plots of TN and TP do not
overlap: the ROC curve tends towards the
point (0,1). This happens because the area
representing the FN or the FP is negligible
in such a classifier histogram.

The Area Under the Curve(AUC) for this near perfect classifier tends to 1
b)If for any classifier, more the TP and FP plots overlap, more the ROC curve falls towards the yellow
line, and less its AUC.
Hence, AUC is also a measure of the performance of the classifier.

You might also like