SVM Model analysis

[nithinth7]

Hi All,

I would like to analyse the models I create with my training data. For Decision Tree we can do this by ploting the tree, for Random Forest/ Tree Ensemble, we have attribute statistics table to get some hints about the models and NN models are very difficult (or nearly impossible) to understand . But I am clueless about SVM models. Is it possible to analyse and understand SVM models?
If possible, how can I do this?

Any help would be highly appreciated!!

Regards,
Nithin

[lisovyi]

Hi @nithinth7,

in general, SVM builds a hyper-surface that separates classes with the largest gap in a multi-dimentional space. This surface is difficult to visualise for 3 dimensions (=variables) and extremely difficult for more or equal than 4 dimensions (=variables). This would work well for 2 input variables only. There is a good introduction to SVM’s in general on Wikipedia.
Having said that, there are different approaches to do model interpretation in a model-agnostic way (for example, by identifying which variables had the dominant effect on the prediction on a row-by-row basis or averaged over a datasample). KNIME 4.0 comes with a reach toolkit of model interpretation algorithms (SHAP and LIME here):
https://hub.knime.com/knime/extensions/org.knime.features.mli/latest/org.knime.mli.node.explain.shap.node.ShapLoopStartNodeFactory
https://hub.knime.com/knime/extensions/org.knime.features.mli/latest/org.knime.mli.node.explain.lime.node.LIMELoopStartNodeFactory

There are also nodes to calculate exact Shapley values, but SHAP does a very good approximation and typically runs faster that the exact calculation.

Cheers,
Misha