Monitoring and Mitigating Calibration Drift in Continuously Updated Evidence-Graded Knowledge Graphs

Key Points

• The aim is to address and mitigate calibration drift in continuously updated knowledge graphs.
• Evaluated CUSUM, EWMA, and KS-test methods for drift detection across four scenarios.
• Compared recalibration strategies including Platt re-fit, isotonic regression, and Bayesian update.
• Part of the Evidence-Graded Knowledge Graphs research programme.
• Identified effective methods for detecting calibration drift.
• Demonstrated that recalibration strategies vary in their effectiveness across scenarios.

Abstract