Drift Config

PSI Drift Configuration

All models that create a PsiDriftProfile will require a PsiDriftConfig object.

from scouter.alert import PsiAlertConfig
from scouter.drift import PsiDriftConfig

PsiDriftConfig(
    name="wine_model",
    space="wine_model",
    version="0.0.1",
    alert_config=PsiAlertConfig()
    categorical_features = ["feature_1", "feature_2"],  # (1)
)

1. To ensure accurate PSI calculations, categorical features must be explicitly specified.

Parameters

Parameter	Type	Description	Example
name	str	The name of the model or dataset being monitored. Defaults to '__missing__' if not provided.	config.name → "wine_model"
space	str	The space where the model or dataset is stored. Defaults to '__missing__' if not provided.	config.space → "wine_model"
version	str	The version of the model or dataset being monitored. Defaults to '0.0.1' if not provided.	config.version → "0.0.1"
alert_config	PsiAlertConfig	Configuration for alerting when drift is detected. Defaults to the default implementation of PsiAlertConfig if not provided.	config.alert_config → *Instance of PsiAlertConfig*
targets	list[str]	List of target features, typically the dependent variable(s).	config.targets → ["churn"]
config_path	Optional[Path]	Path to a pre existing PsiDriftConfig. Defaults to None if not provided	config.config_path → Path("/configs/drift.yaml")
categorical_features	Optional[list[str]]	To ensure accurate PSI calculations, categorical features must be explicitly specified.	config.categorical_features → ["feature_1", "feature_2"]

Properties

Property	Type	Description	Example
name	str	The name of the model or dataset being monitored.	config.name → "wine_model"
space	str	The space where the model or dataset is stored.	config.space → "wine_model"
version	str	The version of the model or dataset being monitored.	config.version → "0.0.1"
feature_map	FeatureMap	When a non-numeric covariate is detected, each unique value is assigned a corresponding numeric value. This mapping is represented by feature_map.	config.feature_map → Instance of FeatureMap
targets	list[str]	List of target features, typically the dependent variable(s).	config.targets → label
alert_config	PsiAlertConfig	Configuration for alerting when drift is detected.	config.alert_config → Instance of PsiAlertConfig
drift_type	DriftType	Type of drift profile.	config.drift_type → DriftType.Psi
categorical_features	Optional[list[str]]	List of categorical features	config.categorical_features → ["feature_1", "feature_2"]

Methods

load_from_json_file() (static method)

Loads a PsiDriftConfig instance from a JSON file.

• Parameters:
  • path (Path): The path to the JSON configuration file. This is required to locate and read the configuration file from disk.
• Returns: A PsiDriftConfig instance.
• Return Type: PsiDriftConfig

---

model_dump_json()

Serializes the PsiDriftConfig instance to a JSON string.

• Parameters: None
• Returns: A JSON string representation of the instance.
• Return Type: str

---

update_config_args()

Updates the configuration of the instance with new values.

• Parameters:
  • space (Optional[str]): The space name, if updating.
  • name (Optional[str]): The new name for the configuration, if provided.
  • version (Optional[str]): The version to set, if specified.
  • targets (Optional[List[str]]): A list of target identifiers, if updating.
  • alert_config (Optional[PsiAlertConfig]): The alert configuration, if provided.
• Returns: None
• Return Type: None

Alert Configuration

An AlertConfig can also be provided to the PsiDriftConfig to specify how you and your team want to be alerted in the event of model drift. The PsiAlertConfig class allows you to configure the alerting mechanism, including the dispatch method (e.g., Slack, OpsGenie), the schedule for drift detection job and even what kind of threshold to use.

from scouter.alert import PsiAlertConfig, SlackDispatchConfig
from scouter.types import CommonCrons

PsiAlertConfig(
    dispatch_config=SlackDispatchConfig(channel='my-team-channel'),
    schedule=CommonCrons.Every6Hours, # (1)
    features_to_monitor=['feature_1', 'feature_2', ...],
    threshold=PsiNormalThreshold() # (2)
)

1. Scouter comes with a set of built-in cron schedules that you can use to configure the schedule for your drift detection job. You can also specify your own custom cron schedule if needed.
2. With the PsiAlertConfig, you can also specify a threshold config that accepts PsiNormalThreshold, PsiChiSquareThreshold, or PsiFixedThreshold.

Parameters

Parameter	Type	Description	Example
dispatch_config	SlackDispatchConfig | OpsGenieDispatchConfig | None	An optional dispatch configuration used to configure how alerts are routed, if None is provided a default internal dispatch type of Console will be used to log alerts to the conosole of the scouter server.	config.dispatch_config -> SlackDispatchConfig()
schedule	str | CommonCrons | None	Schedule to run drift detection job. Defaults to daily at midnigh. You can use the builtin CommonCron options or specify your own custom cron.	config.schedule → CommonCrons.Every6Hours
features_to_monitor	list[str]	List of features to monitor. Defaults to empty list, which means all features	config.features_to_monitor → ['feature_1, feature_2, ...']
threshold	PsiNormalThreshold | PsiChiSquareThreshold | PsiFixedThreshold	The type of threshold to use.	PsiNormalThreshold()

Threshold Types

Out of the box, Scouter provides three types of threshold that can be use with PSI

PsiFixedThreshold

PsiFixedThreshold is the simplest threshold type that allows you to specify a fixed threshold value for drift detection. In a lot of industry settings, this is typically set between 0.10 and 0.25, but you can adjust it based on your specific use case. Note: This is not the most scientific way to detect drift, but it is offered given that most users of PSI are familiar with this approach. However, we recommend using the PsiNormalThreshold or PsiChiSquareThreshold for more robust drift detection.

PsiNormalThreshold

PsiNormalThreshold uses the asymptotic normal distribution of PSI to calculate the threshold for drift detection. This is determined at runtime during drift detection and is based on the over sample size of observed bin data.

The basic premise is that the PSI statistic can be approximated by a normal distribution when there's no drift. Thus, using an observed sample size and pre-defined significance level, we can calculate the critical value for PSI and compare is against the observed PSI value providing us a more dynamic and scientific way to detect drift.

PsiChiSquareThreshold

PsiChiSquareThreshold is similar to PsiNormalThreshold, but uses the Chi-Square distribution to calculate the threshold for drift detection.