Panel data

In panel data the same entity contributes several similar objects: for example, a policy that renews every year. A split of such data by object puts near-duplicate objects of the same entity on both sides. The validation dataset then looks like the training data, early stopping does not trigger, and pruning keeps tables it should drop.

Pass the entity of each object in the groups parameter of fit:

model = TBoostRegressor(objective="poisson")
model.fit(train_data, "ClaimCount", exposure="Exposure", groups="PolicyID")

Whole groups are then assigned to one side of every internal split: the validation objects of early stopping, the bag samples and their out-of-bag objects, and the pruning folds. The groups column is not used as a feature. A grouping in which every group has one object is the same as no grouping.