Drift Detection

Drift Detection

Monitor production data for distribution shifts using PSI and KS tests.

How It Works

Metric	What it detects	Threshold
PSI (Population Stability Index)	Overall distribution shift	> 0.2 = significant drift
KS test (Kolmogorov-Smirnov)	Maximum difference between CDFs	p-value < 0.05 = drift

A feature is flagged as "drifted" if either PSI or KS detects it.

Python API

from scomp_link import DriftDetector

# Fit on training (reference) data
detector = DriftDetector(X_train, psi_threshold=0.2, ks_alpha=0.05)

# Check production data
report = detector.detect(X_production)
# Returns DataFrame: feature, psi, psi_drifted, ks_statistic, p_value, ks_drifted, drifted

summary = detector.summary(report)
# {'total_features': 10, 'drifted_features': 2, 'drift_pct': 20.0, ...}

# Visualize
fig = detector.plot_drift_report(report)             # PSI bar chart
fig = detector.plot_feature_distribution('age', X_production)  # overlay histogram

CLI

scomp-link drift --reference train.csv --current production.csv --threshold 0.2

Integrating with ScompArtifact

# The artifact stores sample data — use it as reference
artifact = ScompArtifact.load("model.scomp")
detector = DriftDetector(artifact.sample_data)
report = detector.detect(new_production_data)