Why are recall for tests and train are diff in each case?

[booramaravind]

Attached are 3 KNIME workflows exported from my workspace with the reset option unchecked. All of them are tuned random forests with same split creation, tree dept, min node leaves, and max models = 100. Why are recall for tests and train are diff in each case? It should be same I think . Which model is best, and why?

Thanks in advance.

Please help with the doubt.
Hotel_Booking_Random_Forest_VP.knwf (1.1 MB)
Hotel_Booking_Random_Forest_VP_2.knwf (1.1 MB)
Hotel_Booking_Cancellation_Prediction_Tunned_Random_Forest_VP.knwf (1.1 MB)
Comparison of performance.docx (53.1 KB)

[ScottF]

Hi @booramaravind -

I didn’t do a deep dive here, but I suspect it’s because you are not holding the seed constant in the Partitioning node. Thus you are actually using a different training split in each workflow.

[Daniel_Weikert]

The “best” model is probably the one which scores highest on your evaluation metric and that metric might depend on your goal
br