multiclass SVM

[William29]

I tried to run an SVM model that I built within SAS viya using 'build model'. However, I received an error indicating that only binary or interval type targets are supported.

In my case, I have approximately 100 classes so I am wondering if there is any easy way that SVM can be used for multi-class (more than 2) classification within SAS viya?

I am aware that this could be done by subsuming all of the classes into just two broader classes, and using binary svm, and then continually repeating the process (of dividing these broader classes into two slightly less broad classes) to classify the observations into each of the 100 classes but I would prefer it if there is a simpler and less time-expensive solution.

[sbxkoenk]

I think this is your only option :

• One-vs-Rest and One-vs-One for Multi-Class Classification
  https://machinelearningmastery.com/one-vs-rest-and-one-vs-one-for-multi-class-classification/
• Paper SAS434-2017
  Methods of Multinomial Classification Using Support Vector Machines
  Ralph Abbey, Taiping He, and Tao Wang, SAS® Institute Inc.
  https://support.sas.com/resources/papers/proceedings17/SAS0434-2017.pdf

Koen

[sbxkoenk]

If you have ~100 or 100+ output labels, I guess you are assigning texts (unstructured data) to topics or so.

If one text can have many topics (like 3 or 5 or 10), you need multi-label models.

Assigning one text to one multinomial output class would be "wrong" in that case.

You can try "extreme multi label text classification".
See here : https://pypi.org/project/extremetext/

See also this :

• Multilabel Classification: An Introduction with Python’s Scikit-Learn
  https://www.kdnuggets.com/2023/08/multilabel-classification-introduction-python-scikitlearn.html
• 1.12. Multiclass and multioutput algorithms
  https://scikit-learn.org/stable/modules/multiclass.html

You can do it through SAS.

Koen