Tidymodels: Ordinal Response and Random Forests


I have been debating the best approach for my data. My research question fits into the classic framework of ordinal regression. I am investigating the most predictive features for "error severity". In short, the inspector is grading the quality of a product produced by someone else. The inspector buckets the errors into 4 categories of "error severity":

  • no error
  • minor error
  • moderate error
  • severe error

I know that ordinal regression fits this scenario. But, I have reason to believe that the relationship with predictors is highly non-linear. So, I want to consider a tree-based algorithm like random forests. The only problem is, I don't think tidymodels has support for ordinal random forests. I've looked into the "ordinalForest" package. But, haven't found much success.

Technically, my research centers around the most severe errors. So, I could recategorize my data into a binary classification. But, I want to get ideas from the community. For example, will multi-level classification work well even though it doesn't capture the inherent order of the response variable?


Hi @MattM ,

Just a quick observation: Folks here are more likely to help with a search for tools that could implement ordinal random forests, but less likely to help sort out the question

This is not say that you won't get replies to both questions here, just that tool-oriented questions usually get more attention. For a more tool-agnostic question like the one above, it might help to also post to either: