'FitPosterior' option for fitcecoc

I am evaluating SVM ('fitcecoc' function) by applying my data 'pm_pareto_12456'. When I set 'FitPosterior' option 'true', I encountered unexpected result described as follows: I execute prediction by using original data. when 'FitPosterior' option is false, the result is same as original classification 'class_array_12456', however, when 'FitPosterior' option is true, some elements of 'predicted_class_12456_true' are different from 'class_array_12456'. I wonder if 'FitPosterior' option could affect the prediction results?

SVM_data.mat contains variables for my evaluation and SVM_Test.m is evaluation script. P.S. my matlab version is 2017a

Answer 1

[Carl]

Hi Yoshihiro. See the documentation here for the 'FitPosterior' option in fitcecoc:

http://www.mathworks.com/help/stats/fitcecoc.html?s_tid=srchtitle#input_argument_d0e285622

It mentions that when that option is set to true, it translates classification scores to posterior probabilities. The model will inherently use different values for its fitting and prediction. As you observed, this may end up with different results.

[Yoshihiro Yamada]

Hi Carl.

According to your advice, I should accept instability of predicted results. Therefore, I tried to change another option to avoid instability and it works well! Thank you for your advice.

[Han zid]

Hi Yoshihiro, I get to same probelem as yours , can you please help me with the option you used to solve this problem.