An override replaces a model’s output for a specified situation when that output is considered unreliable or unsuitable. It can provide a temporary, rule-based response while a lasting model fix is prepared. section 13.4
Use overrides deliberately, and plan for what happens next:
- Replace predictions that fall outside an acceptable range or have too little confidence, for example with a constant value. section 13.4
- For a specific failure reported by a customer, apply a rule-based correction while preparing the scenario for training and a later model release. section 13.4
- Track overrides and plan to address them; untracked rules can be forgotten and become technical debt. Collections of overrides can also inform weak supervision. section 13.4
In an internal prediction system, a manual override function can give business users control over unforeseen or unique circumstances and provide feedback for improving predictions. section 13.5.1
