Skip to main content

databricks.labs.dqx.anomaly.training_strategies

Training strategy pattern for row anomaly detection.

Enables different anomaly detection algorithms through a common interface. Currently implements IsolationForest, but designed for extensibility.

Uses dependency injection for the model registry, enabling:

  • Consistent registration path with EnsembleTrainer
  • Easy mocking/testing
  • Potential for alternative backends

AnomalyTrainingStrategy Objects​

class AnomalyTrainingStrategy(ABC)

Training strategy interface for row anomaly models.

Implement this interface to add new anomaly detection algorithms. Uses dependency injection for the model registry.

__init__​

def __init__(registry: ModelRegistryBase | None = None) -> None

Initialize strategy with optional registry.

Arguments:

  • registry - Model registry to use. Defaults to MLflow/Unity Catalog.

train​

@abstractmethod
def train(train_df: DataFrame, val_df: DataFrame, columns: list[str],
          params: AnomalyParams, model_name: str, *,
          allow_ensemble: bool) -> TrainingResult

Train an anomaly detection model.

Arguments:

  • train_df - Training DataFrame
  • val_df - Validation DataFrame
  • columns - Feature columns to use
  • params - Training parameters
  • model_name - Name for registered model
  • allow_ensemble - Whether to allow ensemble training

Returns:

TrainingResult with model URI, metrics, and metadata

IsolationForestTrainingStrategy Objects​

class IsolationForestTrainingStrategy(AnomalyTrainingStrategy)

IsolationForest training strategy (default).

Uses sklearn's IsolationForest algorithm with optional ensemble training. Both single-model and ensemble paths use the same ModelRegistryBase abstraction.

train​

def train(train_df: DataFrame, val_df: DataFrame, columns: list[str],
          params: AnomalyParams, model_name: str, *,
          allow_ensemble: bool) -> TrainingResult

Train IsolationForest model(s).

If allow_ensemble and params.ensemble_size > 1, trains an ensemble. Otherwise trains a single model using the registry abstraction.