leaf-adjacent models for explainability

[bcjaeger]

this is an idea to make oblique RFs more explainable for a single prediction:

1. pull out the regression coefficients from the node directly above the given obervation's predicted leaf node of each tree
2. aggregate them
3. use the aggregate regression coefficients as a starting value in a regression model fitted to the training set
4. (optional) weight the training set by nearest neighbors of the observation

We would assess the validity of this by measuring correlation between the forest's predictions and the model's predictions. It would necessitate fitting 1 model per prediction, so not computationally great.

[emilyriederer]

Hi @bcjaeger ! Thank you for the awesome package. I think this is a really elegant extension of RFs. I especially love the flexibility in the different parameterizations of the oblique splitters; I'd started out trying treeple in python, but it focuses mostly on a random search of projections per the original paper.

I'm experimenting with a bit of an unusual use case for oblique trees, and for that I'm very interested in being able to extract the exact splitting logic use at each node. I'm commenting here because this sounds similar to your point (1) in your list. Is this currently possible with aorsf? I tried to get a sense of where the projections might be kept and traced the package's train and predict functions; however, as those both defer to C++, I wasn't able to se where these might get stored.

Thank you!

[bcjaeger]

Hi Emily! Thank you for checking out aorsf. =] I think there are some tools that we can use for your experiments, although they aren't as developed as they could be. For example, I think we could use the data stored in an aorsf model to reproduce results that occur while trees are grown. Would it be helpful if I wrote some example code in R for that?

[emilyriederer]

Hi @bcjaeger ! Thanks so much for the fast reply. I hate to ask you to do work, but definitely be very grateful for any psuedocode or hints you might have.

For broader context, I'm imagining a bit of an "off label" use case for aorsf. In industry, I find we often are using features that separately represent the numerators and denominators or ratios that would be more contextually relevant, but often that feature engineering isn't done. I have a theory that in these cases, oblique trees will dramatically outperform normal trees. However, one step further, I'm curious to see if analyzing the splits discovered by oblique trees can "discover" what features could/should be created in regular trees.

So, any information I can figure out how to extract from the trees about the approximate splitting logic used is very useful!