What methods can be used to transform data?

Question

I am solving a binary classification problem with 4 predictor variables. The variables didn't seem to be linearly separable. I have used Neural Networks and Kernel SVM which work and give desired accuracy , but in turn are too complex to interpret and have latency issues.

Are there any transformation like Box-Cox or power transform that I can apply on the data and then use logistic regression/ decision tree for classification. I can sacrifice few % of accuracy to get a simple interpretable model.

What different methods can be used for data transformation?

Answer 1

What methods can be used to transform data?

You can use any function which is defined for your data. If you want to do formal probability theory you can think about it more carefully as a measurable function.

What different methods can be used for data transformation?

This is unbounded in theory and innumerable in practice.

It is largely up to you to figure out which transformations best suite the purpose of the analysis.

Answer 2

Since you are not getting a linearly separable decision boundary and at the same time you don't want complex decision boundaries based on SVM Kernels or neural networks what you can try is build features that tries to capture interaction such as X1*X2 and X1*X2**2 and likewise and include them in your classifier as features. Also you can try tree based classifiers since it captures interaction automatically. Hope this helps ..