Tasks & Targets

Every scenario requires two things: a task that defines the type of prediction, and a target function that tells the model what to learn. Together they turn the foundation model's general understanding into a focused business prediction.

Choosing a Task

The task determines the shape of the model's output. Use the flowchart to find the right task type, then check the comparison table for details.

Task type comparison

Task Type	Details
Binary Classifier	Class: `BinaryClassificationTask` Question: Will something happen — yes or no? Example: Churn, fraud, conversion Target return: `[0]` or `[1]` Output: Probability Default metric: AUROC
Multi-Class Classifier	Class: `MulticlassClassificationTask` Question: Which one of several categories? Example: Favorite brand, plan tier, segment Target return: Normalized scores (sum to 1) Output: Softmax probabilities Default metric: MultipleTargetsRecall Required param: `class_names`
Multilabel Classifier	Class: `MultilabelClassificationTask` Question: Which combination of labels? Example: Cross-sell departments, category affinity Target return: Independent 0/1 flags Output: Sigmoid probabilities Default metric: AUROC Required param: `class_names`
Regression	Class: `RegressionTask` Question: How much? A numeric value. Example: 90-day spend, LTV, days to next visit Target return: Continuous float(s) Output: Predicted value(s) Default metric: RMSE
Recommendation	Class: `RecommendationTask` Question: Which items should be suggested? Example: Next product to buy, content picks Target return: Sketch (weighted item set) Output: Ranked item list Default metric: HitRate@K

For detailed configuration of each task type, see Model Configuration.

Writing a Target Function

The target function is a Python function you write that defines what the model learns to predict. It receives an entity's history, future, attributes, and context, and returns a label matching the task type. See Target Function for the full API.

Python

def churn_target(history: Events, future: Events, entity: Attributes, ctx: dict):
    target_window = timedelta(days=30)
    if has_incomplete_training_window(ctx, target_window):
        return None  # skip entities without enough data

    future = future.interval_from(ctx[SPLIT_TIMESTAMP], target_window)
    return np.array([1 if future["transactions"].count() == 0 else 0],
                    dtype=np.float32)

In This Section

Guide	Description
Target Function	Inputs, time windows, operations on events, and output format in detail
Target Examples	Ready-made examples for each task type
Validation & Best Practices	Validate with `verify_target()`, common mistakes, and performance tips

Once task and target are defined, move on to Model Configuration to set up and run training.