TabularTask

Describes a flat, single-table prediction task for ICLEstimator — no relational schema, no foreign/candidate keys, no DFS. TabularTask supports every task type ICL supports on a flat table, including forecasting, which has no equivalent on the relational (NodeTask) path.

Used together with TabularDataset, which simply wraps a connector and a TabularTask.

Parameters

Parameter	Type	Description	Optional
name	str	Human-readable task name.	No
task_type	TaskType	One of TaskType.BINARY_CLASSIFICATION, MULTICLASS_CLASSIFICATION, REGRESSION, or FORECASTING.	No
label_column	str	Column holding the target labels or values.	No
train_table	str	Path to the training split. Optional — recorded for reference only. ICLEstimator never uses it as context; there is no training loop, so it has no other purpose here.	Yes
validation_table	str	Path to the validation split, required by score() (or pass val_table= at call time).	Yes
test_table	str	Path to the test split, required by predict() (or pass test_table= at call time).	Yes
context_table	str	Table used as the zero-shot ICL context. Required for predict()/score() — must be set explicitly.	Yes (required at call time)
time_column	str	Datetime column. Required when task_type=FORECASTING; also used by the "most_recent"/"mixed" sampling strategies for classification/regression.	Yes / required for forecasting
item_id_column	str	Series partition key for panel forecasting (multiple entities, one series each). Omit (None) for single-entity forecasting.	Yes
prediction_length	int	Steps ahead to forecast. Required when task_type=FORECASTING; must be >= 1.	Required for forecasting
max_context_length	int	Max history length (in time steps) the forecaster looks back over. Required when task_type=FORECASTING; must be >= 1.	Required for forecasting
evaluation_metric	EvaluationMetric	Metric to compute in score(). Auto-selected when None. Must be a valid metric for task_type.	Yes
column_dtypes	Dict[str, str]	Optional dtype overrides applied before featurization.	Yes

FORECASTING requires more than the other task types

When task_type=TaskType.FORECASTING, prediction_length, time_column, and max_context_length are all required — the constructor raises a ValueError immediately if any is missing. No other task type has this requirement.

Returns

An instance of the TabularTask class.

Example — classification

from relationalai_predictive import TabularTask, TaskType

task = TabularTask(
    name="churn",
    task_type=TaskType.BINARY_CLASSIFICATION,
    label_column="churned",
    context_table="DATABASE.SCHEMA.CHURN_CONTEXT",
    test_table="DATABASE.SCHEMA.CHURN_TEST",
    validation_table="DATABASE.SCHEMA.CHURN_VALIDATION",
)

Example — forecasting

from relationalai_predictive import TabularTask, TaskType

task = TabularTask(
    name="sales_forecast",
    task_type=TaskType.FORECASTING,
    label_column="sales",
    context_table="DATABASE.SCHEMA.SALES_CONTEXT",
    test_table="DATABASE.SCHEMA.SALES_TEST",
    time_column="date",
    item_id_column="store_id",
    prediction_length=12,
    max_context_length=64,
)