How to interpret xerror of rpart (method="class") when losses are assigned?

[A-Pai]

How to interpret xerror of rpart (method="class") when losses are assigned

Strangely initial xerror is extremely small and it is 0.1000 (I see this is 1/loss of class 0), then xerror increases to 2.41, 2.77 and so on. Somehow xerror at the initial node is much much better than rel error!!

[hmeeks0212]

Have you figured out what to do yet? Because of the loss matrix, the 1-SE rule doesn't apply. I don't how how to pick the trees in this case.

[bethatkinson]

Nope - I don't know what to recommend. Good luck figuring it out. From: hmeeks0212 ***@***.***> Sent: Wednesday, October 23, 2024 10:55 PM To: bethatkinson/rpart ***@***.***> Cc: Subscribed ***@***.***> Subject: [EXTERNAL] Re: [bethatkinson/rpart] How to interpret xerror of rpart (method="class") when losses are assigned? (Issue #65) Have you figured out what to do yet? Because of the loss matrix, the 1-SE rule doesn't apply. I don't how how to pick the trees in this case. - Reply to this email directly, view it on GitHub<#65 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ACWQG56GCOS4VYU7PTZ7PXDZ5BVQ7AVCNFSM6AAAAABQQEZBWWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIMZUGE4TMOJWGE>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.******@***.***>>

[hmeeks0212]

Ah, that's OK. Thank you so much for responding so quickly!

First of all, I would like to thank you for all the extensive work that you've done to maintain rpart. When I asked my previous question, I thought I was communicating with Apai!

I'm not sure if this would be of help to you. But an additional problem that I've noticed is that including the loss matrix would impact not only the xerror, but also the variable of importance. The rpart manual stated that the overall measures of variable importance are scaled to sum to 100. This is not always the case when including the loss matrix.

Additionally, I have been trying to replicate the trees created in rpart and Minitab. I've read the manuals for both rpart and Minitab Predictive Analytics software. My understanding from studying the two manuals is that both rpart and Minitab use the same method. However, Minitab applied additional logic (that I couldn't find online) to calculate xerror and xstd when loss matrix is applied. I thought I would let you know just in case you can find or have access to the underlying logic that Minitab uses.

I hope that would help with any future update for rpart.

Best,
M.

[bethatkinson]

I have no idea - I haven't used minitab in over 30 years From: hmeeks0212 ***@***.***> Sent: Friday, October 25, 2024 12:42 AM To: bethatkinson/rpart ***@***.***> Cc: Atkinson, Beth J., M.S. ***@***.***>; Comment ***@***.***> Subject: [EXTERNAL] Re: [bethatkinson/rpart] How to interpret xerror of rpart (method="class") when losses are assigned? (Issue #65) Ah, that's OK. Thank you so much for responding so quickly! First of all, I would like to thank you for all the extensive work that you've done to maintain rpart. When I asked my previous question, I thought I was communicating with Apai! I'm not sure if this would be of help to you. But I have been trying to replicate the trees created in rpart and Minitab. I've read the manuals for both rpart and Minitab Predictive Analytics software. My understanding from studying the two manuals is that both rpart and Minitab use the same method. However, Minitab applied additional logic (that I couldn't find online) to calculate xerror and xstd when loss matrix is applied. I thought I would let you know just in case you can find or have access to the underlying logic that Minitab uses. - Reply to this email directly, view it on GitHub<#65 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ACWQG527S7HO22XFN754VCLZ5HK4BAVCNFSM6AAAAABQQEZBWWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIMZWHA4TANJXGM>. You are receiving this because you commented.Message ID: ***@***.******@***.***>>