Reliable ML: Model Robustness After Production Rollout

Dmitry Kolodezev27 May 2023

Problem

A model that passed acceptance starts losing money after release, and often nobody notices. A service runs on the wrong port, the data distribution shifts, or marketing renames a classifier in the CRM that the model relies on. The model is one step in a business process, often built by an outsourcer who has left. The owner pays unbudgeted operating costs, and efficiency gains shrink month after month. When labels mature slowly, quality cannot be tracked directly: a bank cannot issue bad loans for 60 days to find out the model has degraded.

Example from the article: in a CRM system, marketing constantly renamed classifiers the model relied on, so the model was constantly wrong.

Solution in brief

Choose monitoring by label maturation speed. With fast labels, track quality against a service-level objective and roll back on a breach. With slow labels, watch proxy metrics, the prediction histogram, input data and a second model that predicts the first one's errors. Before launch, add fallbacks and anomaly handling, and weigh retraining cost against the loss from staying put.

Questions this case answers
  • How to monitor a deployed prediction model when true outcomes arrive weeks or months after the prediction
  • How to decide when a deployed prediction model must be retrained given its degradation pattern and the cost of retraining
  • How to monitor a production model when its labels arrive weeks or months later
  • How to estimate how quickly a deployed model degrades and when to retrain

Tools and methods

  • SLO / SLI / SLA — target quality, its windowed average, the promise to the business
  • Proxy metrics and prediction histograms — fast stand-ins for slow labels
  • Evidently AI — data, label and concept drift
  • NannyML — residual (error) modeling
  • Interleaving — comparing a retrained model with the old one
  • binsreg — mean-target plot with confidence intervals
Continue reading — it's free

Google or a one-time email code. No password.