search

kapelner / bartMachine

An R-Java Bayesian Additive Regression Trees implementation
MIT License
62 stars 27 forks source link

DAG as a way to visualise decisions of the trained BART tree #51

Closed sunPi closed 1 year ago

sunPi commented 1 year ago

First of all, thank you for this implementation it works really well in R and is easy to use as well as efficient.

There are a lot of implemented functions that allow the user to measure the performance of a BART model. Other metrics are normally implemented by the user (e.g. correlation or R2 etc.). What about explaining the model? I don't know if there is an implemented way to draw the final decision tree that was made by the Bart machine.

For example, I am looking for a way to generate a DAG graph from the decisions made by the final BART tree. This way we could calculate the importance/contribution of each feature to the final model performance, possibly implementing SHAP values as well.

In short, is there a way to draw the BART tree as a DAG or any other way to measure feature contribution?

kapelner commented 1 year ago

Bart is a sum of trees model. If you want one tree, use the rpart package which also has nice visuals. If you want to "explain" the Bart model, good luck. Nobody knows how to explain machine learning. Check out Stephen Wolfram's testimony before the United States congress on YouTube.

On Thu, Aug 31, 2023, 08:47 sunPi @.***> wrote:

First of all, thank you for this implementation it works really well in R and is easy to use as well as efficient.

There are a lot of implemented functions that allow the user to measure the performance of a BART model. Other metrics are normally implemented by the user (e.g. correlation or R2 etc.). What about explaining the model? I don't know if there is an implemented way to draw the final decision tree that was made by the Bart machine.

For example, I am looking for a way to generate a DAG graph from the decisions made by the final BART tree. This way we could calculate the importance/contribution of each feature to the final model performance, possibly implementing SHAP values as well.

In short, is there a way to draw the BART tree as a DAG or any other way to measure feature contribution?

— Reply to this email directly, view it on GitHub https://github.com/kapelner/bartMachine/issues/51, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAFAV6GRMVYWN2UHR6GT5MDXYCBWBANCNFSM6AAAAAA4GAVZMM . You are receiving this because you are subscribed to this thread.Message ID: @.***>