Categorical split support for SHAP

[trivialfis]

Depends on rapidsai/gputreeshap#26 .

[trivialfis]

Please do not merge this one until I can replace the submodule back to rapidsai.

[talieh-tabatabaei]

How Shap values for each one-hot-encoded category are aggregated to generate the final value for the categorical variables?

[oOTWK]

@talieh-tabatabaei When enable_categorical=True is enabled, it does not compute Shap values for each one-hot-encoded variable. To enable this for Shap computation, the tree model should be trained with this enabled. In this case, there are no one-hot-encoded variables anymore, it's just one categorical variable itself. So decision variable in the tree model is the categorical variable.

For example, let's say we have Fruit category and the values are Apple, Orange, and Melon. A node of tree could have Fruit==Orange. In one-hot-encoded version, it is equivalent to Is_Orange==1 but in this case, Is_Orange, Is_Apple, and Is_Melon are different variables. In TreeShap, these variables do not invoke Unwind to each other. On the other hand, if we just use the categorical variable Fruit, Fruit==Orange and Fruit==Apple and Fruit==Melon all use the same variable Fruit. This does invoke Unwind if there is more than one Fruit variable appearing on the current path. As a result, it will only compute Shap value for Fruit variable.