model.analyze() function Partial Dependence Plot axis scaled incorrectly

[physics2001]

I was running the GradientBoostedDecisionTreesLearner on a classification dataset with some binary variables, and the plots for binary variables are all between 0.25 and 0.75 instead of 0 and 1 as shown below. Instead of zoomed in like this, the PDP should always display the change from minimum value to maximum value. For the method call, I just used the following: model.analyze(df, sampling=1)

[rstz]

Thank you for letting us know. The root cause is that the PDP of numerical values is created by binning the attribute values (x-axis) and then averaging the predictions over each bin (y-axis). If the value is binary but numerical, the bins will be [0, 0.5], centered around 0.25, and [0.5, 1], centered around 0.75.

The workaround would be to declare the corresponding feature as categorical. I'll see if there's a reasonably-scoped fix we can do even if the feature is declared numerical.

[physics2001]

Thank you, I realized that as well. I switched them to string, and it helped resolve the issue. It would be nice to improve YDF to support it since some categorical features contain discrete integers.