1
$\begingroup$

I'm getting different results between R and Python for a classification problem using ordinal categorical features.

I would like to ask you what do you think is the best classification algorithm in Python to use for ordinal categorical data?

In R this seems to be some Machine Learning algorithm based on "ctree". In Python, TreeDecisionClassifier, for example, is based on "CART". I didn't find any "ctree" based Machine Learning Algorithm reference for Python.

My dataset is survey results. All my features have values from 1 to 5 (bad to good), and the target variable is the NPS which takes values from 1, 2 and 3 (detractor, neutral and promoter).

Thank you.

$\endgroup$
2
  • $\begingroup$ Do you have any more information, such as what type data set you are working with? Or is the question you want the ctree equivalent for Python? $\endgroup$
    – M__
    May 17, 2023 at 22:19
  • 1
    $\begingroup$ Both. I updated it. $\endgroup$
    – user140259
    May 17, 2023 at 22:34

1 Answer 1

2
$\begingroup$

ctree or a conditional tree is a non-parametric decision tree. TreeDecisionClassifier in Python is the closest, but there isn't a direct equivalent of a conditional tree in Python's sci-kit learn, which appears the context of the question. It may be possible to modify the scikit learn's decision tree to be a conditional tree - in theory a conditional tree is a subset of a decision tree - but I don't know whether this has been achieved.

Purely personally I wouldn't use sci-kit learn's decision tree because it is prone to overtraining. If I received a manuscript singly using a decision tree, I would return it requesting alternative models to at least supplement the decision tree analysis and request checks for overtraining. Thus in this regard it doesn't have a good reputation. A random forest model would be used instead. ctree is cool.

So to answer your question directly I would not replace a conditional tree with scikit learn's decision tree. If you do then you need to check for overfitting (IMO). However, there was talk of implementing a conditional tree in sci-kit learn.

A random forest is an averaged conditional tree, thus I personally would see it as the preferred Python model for replacing a ctree, but its not the same thing.


What would I do? Well if you are happy with ctree and cforest then I would use sci-kit learn's random forest.

Beyond that I don't know enough about your data to make a call (I do evolution). Random forests and conditional trees are good with sparse data, so possibly if there are lots of questions that have been skipped (that might be relevant).

Random forests don't deal with imbalance very well. I don't know your targets, but I could imagine the outcomes could be imbalanced because at a wild guess there are more neutrals then any other category.

xgboost, or extreme gradient boosting, is trendy. It will replace random forests (IMO), especially to overcome imbalance and copes with high variance. On the surface its just as easy to use. The parameterisation can be computationally demanding, but there's short-cuts and there's an interaction analysis for examining the features of the survey. This would be particularly true if you want to improve your survey. However its more complex to grasp all the caveats.

I know about surveys in clinical data and getting a good survey is not trivial. The survey and even the investigator can generate bias, but I suspect it will also depend on the morbidities of the patients. In new clinical surveys a model survey is used that is known to be robust. Its definitely not an easy area of investigation.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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