I'm working on a project that classifies conversation success vs failure based on loads of categorical variables about external interactions. I have already concluded logistic regression would not fit well here (due to messy data woes) and I'm going with decision trees.

The findings must explain what variables lead to successful outcomes and my question is, which decision tree is the best to use in R? I'm between CART, C5.0, and random forests. I know using CART is great for visualization and interpretability but I'm worried about non-robust findings (could change with new data). I know random forests are much less interpretable, but more robust. I know less about C5.0 but that is uses information gain in place of a Gini metric for splitting.

Which is best? Can I show visualizations of C5.0 or any other decision tree not mentioned?


What if you trained a non-interpretable model such as a random forest to reduce variance and achieve your intended robustness, but then trained a surrogate decision tree around your random forest's predictions to get a rough sense of what the model actually does?

https://github.com/h2oai/mli-resources has links to some good resources on machine learning interpretability - https://github.com/h2oai/mli-resources/blob/master/notebooks/dt_surrogate.ipynb is a notebook that shows how to build a surrogate decision tree for interpretability.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.