Logged Bandit from a DataFrame

This guide shows how to build a LoggedBanditDataset from a pandas DataFrame.

Expected columns

You need:

• context features (one or more columns)
• action column with integer action indices
• reward column
• optional behavior_prob column for logged propensities

Example

from crl.data.datasets import LoggedBanditDataset

# df columns: ["x1", "x2", "action", "reward", "behavior_prob"]
bandit = LoggedBanditDataset.from_dataframe(
    df,
    context_columns=["x1", "x2"],
    action_column="action",
    reward_column="reward",
    behavior_prob_column="behavior_prob",
)

Common pitfalls

• Action indices are not 0..K-1: remap your actions before creating the dataset.
• Propensities missing: IS/WIS require behavior_action_probs. See Behavior Propensities Missing.
• Contexts are mixed types: cast to numeric or one-hot encode categorical features.