Skip to content
Scouter Scouter

PSI quickstart

Population Stability Index (PSI) is one of the most common ways to monitor ML models in production. The following sections will walk you through:

  • Setting up a PSI drift profile
  • Configuring real-time notifications for model drift detection

Creating a Drift Profile

Section titled “Creating a Drift Profile”

To detect model drift, we first need to create a drift profile using your baseline dataset, this is typically done at the time of training your model.

from scouter.alert import PsiAlertConfig, SlackDispatchConfig
from scouter.client import ScouterClient
from scouter.drift import Drifter, PsiDriftConfig
from scouter.types import CommonCrons
from sklearn import datasets


# Prepare data
X, y = datasets.load_wine(return_X_y=True, as_frame=True)


# Drifter class for creating drift profiles
scouter = Drifter()


# Specify the alert configuration
alert_config = PsiAlertConfig(
    features_to_monitor=["malic_acid", "total_phenols", "color_intensity"], # Defaults to all features if left empty
    schedule=CommonCrons.EveryDay, # Run drift detection job once daily
    dispatch_config=SlackDispatchConfig(channel="test_channel"), # Notify my team Slack channel if drift is detected
    # psi_threshold=0.25  # (default) adjust if needed
)


# Create drift config
psi_config = PsiDriftConfig(
    name="wine_model",
    space="wine_model",
    version="0.0.1",
    alert_config=alert_config
)


# Create the drift profile
psi_profile = scouter.create_drift_profile(X, psi_config) # (1)


# Register your profile with scouter server
client = ScouterClient()


# set_active must be set to True if you want scouter server to run the drift detection job
# You can always set this later
client.register_profile(profile=psi_profile, set_active=True)
  1. The data you pass to create_drift_profile should be the baseline dataset used to train your model. It is recommended to use a representative sample of your training data. It is also expected that the data is free of missing values, nans and infinities, or else the drift profile creation may fail or give unexpected results.