predictorImportance changes between releases

2 views (last 30 days)
I've been debugging an old bagged ensemble tree model and I noticed that when I run the model using MATLAB R2019a, R2017b, and R2021a, I get different results in the ordering and values of the important features using the predictorImportance function. The results between the R2017b and R2021a versions are similar (and close to what I'd expect to be the important features), but the R2019a results are quite different. All I've changed was the version, and printed out the variables, template tree, and model to check. I've looked through the old release notes and didn't see any mention of changes made to this function. I'm wondering if there was another change that influenced this function or if anyone else has encountered this issue?

Answers (1)

the cyclist
the cyclist on 15 Jul 2022
Are the ML model results themselves identical?
Off the top of my head, I don't know enough about the algorithm and possible changes, but one easy thing to try would be to set the random number generator seed at the start of your code, e.g.
rng default
and see if you then get the same results. This is a common cause of differences.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!