Difference fitcecoc(<chosen model template>) and individidual multiclass model classifier (fitcnb, fitcknn, fitcdiscr, fitctree)

Hello dear Matlab-Community,

I am wondering which is difference of fitting a multiclass classifier through "fitcecoc" or the individual model commands as shown above in the summary.

For example, is there a difference in the following commands?

%

mdl = fitcdiscr(X,Y,"OptimizeHyperparameters","all");

%

and

%

tnb = templateDiscriminant;

mdl = fitcecoc(X,Y,'Learners',tnb,"OptimizeHyperparameters","all");

%

I am looking foward to you reply!

Thank you:)

Answer 1

[Vaibhav]

Hi Denys,

It is my understanding that the clarification is requested regarding the distinction between employing “fitcecoc” for fitting a multiclass classifier and using individual model commands like “fitcdiscr”.

Please find the difference below:

Approach:

• fitcdiscr: Trains a single multiclass classifier directly, without decomposing the problem into binary subproblems.
• fitcecoc: Uses the Error Correcting Output Codes (ECOC) approach, decomposing the multiclass problem into multiple binary subproblems

Number of Classifiers:

• fitcdiscr: Trains a single multiclass classifier.
• fitcecoc: Trains multiple binary classifiers, one for each binary subproblem.

Classification Strategy:

• fitcdiscr: Uses the specified discriminant analysis algorithm to directly solve the multiclass problem.
• fitcecoc: Solves the multiclass problem by combining the predictions of the binary classifiers trained for each subproblem.

Flexibility:

• fitcdiscr: Provides a simpler and more straightforward approach when you want to use a specific discriminant analysis algorithm for multiclass classification.
• fitcecoc: Offers more flexibility by allowing you to choose different binary classifiers and leverage their strengths for multiclass classification.

Performance:

• fitcdiscr: Performance depends on the discriminant analysis algorithm chosen and its suitability for the problem.
• fitcecoc: Performance can be influenced by the choice of binary classifiers and their ability to handle the binary subproblems effectively.

Interpretability:

• fitcdiscr: Provides direct interpretability of the multiclass classification results.
• fitcecoc: Requires interpreting the results of multiple binary classifiers to understand the multiclass classification outcome.

In summary, the main difference is that `fitcdiscr` trains a single multiclass classifier, while `fitcecoc` decomposes the problem into binary subproblems and trains multiple binary classifiers. This allows `fitcecoc` to potentially provide better performance by leveraging the strengths of individual binary classifiers. However, `fitcdiscr` is simpler and more straightforward if you want to train a single multiclass classifier using a specific discriminant analysis algorithm.

Please refer to the following documentations for more information:

https://in.mathworks.com/help/stats/fitcdiscr.html

https://in.mathworks.com/help/stats/fitcecoc.html

https://in.mathworks.com/help/stats/fitcecoc.html#bufm0tv:~:text=Error%2DCorrecting-,Output%20Codes%20Model,-An%20error%2Dcorrecting

Hope this help!

Regards,

Vaibhav