Talk:Binary classification
![]() | This article is rated Start-class on Wikipedia's content assessment scale. It is of interest to the following WikiProjects: | ||||||||||||||||||||
|
specificity and sensitivity
editspecificity is defined as TN/(TN+FP) and sensitivity is defined as TP/(TP+FN). In this article the definitions given are incorrect (specificity = TN/(TN+FN), sensitivity = TP/(TP+FP)). 99.246.20.224 (talk) 21:28, 20 December 2009 (UTC)
I agree about the discrepancy. The statement "Specificity (TNR) is the proportion of people that are actually negative (TN) of all the people that tested negative (TN+FN)" should read "Specificity (TNR) is the proportion of people that are tested negative (TN) of all the people who are actually healthy (TN+FP)."Toahi (talk) 09:34, 27 February 2010 (UTC)
This correction will be consistent with the follow-up statement 'As with sensitivity ... given that the patient is not sick', which are those people who are actually healthy, or (TN+FP), regardless of any test, and *not* those whose TEST results label them as not sick, or (TN+FN).Toahi (talk) 09:34, 27 , February 2010 (UTC)
The following sentence 'With higher specificity, fewer healthy people are labeled as sick (or, in the factory case, the less money the factory loses by discarding good products instead of selling them)." again corresponds to TN/(TN+FP), where FP= healthy (Positive) people are labeled as sick (False test result). Toahi (talk) 08:38, 27 February 2010 (UTC)
Similarly, the statement "Sensitivity (TPR) is the proportion of people that are actually positive (TP) of all the people that tested positive (TP+FP)" should read " Sensitivity (TPR) is the proportion of people that are tested positive (TP) of all the people that are truly positive (TP+FN)". All the follow-up statements will then be consistent.Toahi (talk) 09:47, 27 February 2010 (UTC)
If sickness is the condition to be detected with least error possible, then 'Specifity' is a measure of how well healthy people are detected, and 'Sensitivity' is a measure of how well sick people are detectedToahi (talk) 09:47, 27 February 2010 (UTC)
Why is this large section on sensitivity and specificity even in this article? Its totally unnecessary, if you wanted to know about sensitivity and specificity you'd go to the article about sensitivity and specificity. — Preceding unsigned comment added by 109.76.224.135 (talk) 17:05, 5 May 2012 (UTC)
TP+FN+TN+FP
editEnglish is not my native language, so I'm very grateful for help with grammar and spelling. // Janka
This sentence in "Evaluation of binary classifiers" might not be correct:
Thus, the number of true positives, false negatives, true negatives, and false positives add up to 100% of the set.
In my opinion TP and FN add up to 100% as FP and TN do as well. //--83.171.184.232 (talk) 22:33, 6 January 2008 (UTC)
- Well a TP is a positive... one that has been accurately predicted as positive. A FN is a positive... one that has inaccurately been predicted as a negative. Together, they make up the set of positive elements. This, however, is only 100% if there are no negative elements.
- Another way of looking at this is as follows: imagine you are doing pregnancy tests. The set of women who are pregnant and are tested as such would be your TP, and then women who are pregnant and who did not show up as such would be your FN. However, this would not necessarily make up 100% of the women tested, as there could be women who are indeed not pregnant. This would be your set of negative elements, and the accuracy of the tests would determine whether they are TN or FP. The total set of women who are pregnant (the positives, regardless of prediction) and those who aren't (the negatives) would be 100% of the women. Hope this helps... WDavis1911 (talk) 16:54, 10 June 2008 (UTC)
Illustration
editI think this graphics is extremely confusing. Even though it seems cute and simple, it has so many annotated arrows compared to Precision and Recall, while having even less explanation. Either way, it is still unintelligible. In Type I and type II errors, the simple matrix of 'Test Result .vs. Actual Condition' illustrates all four possibilities immediately in a familiar matrix notation. The current graphics does not help illustrate the principle, as the reader will be busy deciphering the graphics instead. Toahi (talk) 09:00, 27 February 2010 (UTC)
Trade offs by varying discrimination threshold
editShould the article not discuss trade offs by varying the discrimination threshold. eg the medical test example may choose some specific value of a biomarker as a threshold for making a positive prediction ? Varying the threshold trades off false positives against false negatives. - Rod57 (talk) 18:40, 5 May 2011 (UTC)
Focus of article
editThis is a really poor article. Apart from the last section "Converting continuous values to binary" which is vaguely related, there's nothing in here about *binary classification*. How do you do binary classification? What are the established methods? Statistical methods? Machine learning methods? The article is supposed to be about binary classification not hypothesis testing and sensitivity and sensitivity! — Preceding unsigned comment added by 109.76.224.135 (talk) 17:15, 5 May 2012 (UTC)