concept-drift-analysis

Diagnose concept drift by correlating per-slice AUC and F1 drops with ground-truth labels to distinguish data drift from model degradation. This analysis cross-references performance reports with drift reports to pinpoint the root cause.

Data drift indicates shifting input distributions, whereas model degradation signifies weakened predictive power. Cross-referencing sliced performance metrics with ground-truth signals helps distinguish between the two to determine the correct root cause.

Investigate a sliced AUC drop by reading performance.json and baseline_metrics.json, then correlating those metrics with ground-truth labels. Cross-referencing an optional drift_report.json provides a clear root cause analysis and actionable next steps.

A separate drift report is optional for root-cause analysis. You can diagnose concept drift using performance.json and baseline_metrics.json, but cross-referencing an optional drift_report.json improves the accuracy of distinguishing data drift from model degradation.

After identifying concept drift, recommended next steps involve targeted remediation actions such as model retraining or implementing data-quality fixes. This process yields traceable decision records for your MLOps pipeline.

Baseline metrics from baseline_metrics.json are required to evaluate concept drift. Comparing current sliced AUC and F1 performance against these baseline metrics is essential to detect drops and determine the root cause.