File Exchange

image thumbnail

ROC Curve

version 3.1 (3.55 KB) by Víctor Martínez-Cagigal

Víctor Martínez-Cagigal (view profile)
MATLAB function which performs a ROC curve of two-class data.
4.8
17 Ratings

96 Downloads

Updated 13 Dec 2018

View License

This function calculates the Receiver Operating Characteristic curve, which represents the 1-specificity and sensitivity of two classes of data, (i.e., class_1 and class_2).

The function also returns all the needed quantitative parameters: threshold position, distance to the optimum point, sensitivity, specificity, accuracy, area under curve (AROC), positive and negative predicted values (PPV, NPV), false negative and positive rates (FNR, FPR), false discovery rate (FDR), false omission rate (FOR), F1 score, Matthews correlation coefficient (MCC), Informedness (BM) and Markedness; as well as the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

Example of use:
class_1 = 0.5*randn(100,1);
class_2 = 0.5+0.5*randn(100,1);
roc_curve(class_1, class_2);

Cite As

Víctor Martínez-Cagigal (2019). ROC Curve (https://www.mathworks.com/matlabcentral/fileexchange/52442-roc-curve), MATLAB Central File Exchange. Retrieved .

Comments and Ratings (28)

Sven Månsson

Sven Månsson (view profile)

Knut Andre Skaug

Knut Andre Skaug (view profile)

Víctor Martínez-Cagigal

Víctor Martínez-Cagigal (view profile)

Hi @Knut,

First of all, if one class has values that are always above 0, and the other class has always values under 0, then the ROC curve will be perfect (reaching AROC=1), just because you can discriminate between 2 classes perfectly by putting a theshold T=0.

With regard to the manner in which I compute the thresholds... I just take all the values of both classes and merge them. Then, I sort them and I take the difference between those points. Finally, I sum the half of that difference. Thus, I ended getting points in the middle of every single sample, therefore having the optimal number of threshold values between each single sample.

If you have more questions, please write me an e-mail. An please, if you find this code useful give me a rating, it helps me a lot. Regards,

Knut Andre Skaug

Knut Andre Skaug (view profile)

Hi Victor

Great function!

I don't understand how each threshold values is calculated.
Threshold is:
(difference between all values divided by 2) + all values
Why is that? Could you refer me something to read up on?
The two distributions I am testing has a pre-specified cut-off at zero. So all cases above zero is a positive, all cases below zero is a negative.
The way it is computed in the function produces some false negatives of which there really are none.

Marta Jimenez

Marta Jimenez (view profile)

Amazing work, thank you!!

Bryan Wang

Bryan Wang (view profile)

Vasia Papoutsaki

Vasia Papoutsaki (view profile)

Excellent! Thank you!

sharar

sharar (view profile)

Thank you so much for your reply.

Keshav Kundassery

Keshav Kundassery (view profile)

Perfect, thank you!

Víctor Martínez-Cagigal

Víctor Martínez-Cagigal (view profile)

Hi sharar and Keshav,

Results do not change, just your interpretation of them. This is a basic concept of ROC Curves. You can visualize it by inputting:

class_1 = 0.5*randn(100,1);
class_2 = 0.5+0.5*randn(100,1);
figure(1);
roc_curve(class_1, class_2);
figure(2);
roc_curve(class_2, class_1);

The AUC of the first attempt is 0.7407, while the AUC of the second attempt is 0.2593. Are the results different? No, just the inverse (note that 0.2593 = 1 - 0.7107). That is because you are swapping the positive and negative classes... Just as a recommendation, if any AUC is less than 0.5, that is because you should swap the classes.

sharar

sharar (view profile)

Hi Victor,
Thank you for reply. Just as Keshav has asked, I have not either understood the reason why AUC differs when the classes are swapped.

Keshav Kundassery

Keshav Kundassery (view profile)

Hi Victor,
Why do I get different results when I swap class_1 and class_2?

Víctor Martínez-Cagigal

Víctor Martínez-Cagigal (view profile)

Hi @sharar, the script will still work if the classes are not equal. I solved that issue in the last version.

sharar

sharar (view profile)

Thank you so much for the code. Just one question. What if the size of two classes are not equal?

stilloboy

stilloboy (view profile)

AleCon

AleCon (view profile)

Chang hsiung

Chang hsiung (view profile)

Ahmed iqbal

Ahmed iqbal (view profile)

Excellent code ! Thank you very much !

Monica Esparza

Monica Esparza (view profile)

THANK YOU

Cheer

Cheer (view profile)

thanks

Estefanía Barbero Aguirre

Estefanía Barbero Aguirre (view profile)

Te adoro

Tasnuba Siddiqui

Tasnuba Siddiqui (view profile)

Mariem Slim

Mariem Slim (view profile)

Zhenxi Song

Zhenxi Song (view profile)

Machine Learning Enthusiast

Machine Learning Enthusiast (view profile)

Can you please update this to three classes? or multi class?

Fernando Vaquerizo

Fernando Vaquerizo (view profile)

shahab i

shahab i (view profile)

nice,tnx

Muhammad Arif

Muhammad Arif (view profile)

Hi Victor,

I really thank you for this amazing post! thank you!

Updates

3.1

All parameters are now printed in the CMD.
3.0

The function now outputs more parameters.
2.1.0.0

Classes are now indicated separately.
2.0.0.0

Different sizes in class_1 and class_2 are now allowed.
1.1.0.0

Fixed a bug in the output data.
MATLAB Release Compatibility
Created with R2014a
Compatible with any release
Platform Compatibility
Windows macOS Linux