AI Glossary / Governance, ethics and safety / Intermediate
Post-processing
Definition
Post-processing is the adjustment or transformation of a model's output after inference. It can convert scores into decisions, apply output constraints or prepare results for use without changing the original model's learned parameters.
What it means
A model rarely operates alone. Its raw output may be a score, a ranked list or a set of candidate objects. A later stage decides how that output becomes something the user sees or a downstream process acts upon. That stage is part of the deployed system's behavior.
For example, a classifier may assign a score of 0.62 while an application uses a threshold to decide whether to flag the case. Changing the threshold changes the decision even when the model and its score remain unchanged. In fairness work, post-processing can adjust decisions to pursue a specified group-level criterion.
The term is broader than fairness mitigation. It also covers many output transformations in other settings. It is distinct from post-hoc explanation, which seeks to explain behavior; post-processing changes how outputs are represented or used.
How it works
Identify the output produced by the model and the transformation applied afterwards. The transformation might use a fixed rule, a separately fitted mapping or another optimization procedure. Specify its inputs and ensure the deployment can supply them reliably.
If the transformation is selected using data, separate fitting or selection from final evaluation. Repeatedly adjusting thresholds against the final test set can make performance look better than it will be on fresh cases. Assess the combined model-and-post-processing pipeline, including relevant costs, error rates and operational constraints.
For fairness-related adjustments, define the criterion and affected groups explicitly. Improving one measure can change other measures and the distribution of errors. Some methods need group attributes when making decisions, raising practical and context-specific governance questions. Results achieved on validation data may not hold after population change. Version and monitor the transformation as carefully as the model it follows.
Three practical examples
- A support classifier converts urgency scores into a review queue using an operationally chosen threshold.
- An image detector removes overlapping candidate boxes before presenting the detected objects to a user.
- A team evaluates a decision-threshold adjustment intended to reduce a specified disparity in true-positive rates, while also checking false-positive rates.
Think of it this way
Think of a weather forecast followed by a venue's rule for opening an outdoor area. The forecast supplies an estimate; the rule converts that estimate into a particular action. Changing the rule can change the action without changing the forecast.
The analogy does not imply that the forecast is correct or that any chosen rule is justified. It illustrates why the full decision process includes both prediction and the transformation applied afterwards.
Worked example
The same scores, different decisions
An invented model assigns six cases scores of 0.20, 0.40, 0.55, 0.65, 0.80 and 0.90. With a rule that flags scores at or above 0.50, four cases are flagged. Raising the threshold to 0.70 flags only two.
No model parameter or score has changed. The difference comes entirely from the post-processing rule. Whether the second rule is better cannot be decided from these scores alone: the team needs outcomes, the consequences of missed and unnecessary flags, and the capacity of the review process.
If the threshold is chosen because it happens to perform best on these six cases, its result still needs evaluation on independent appropriate data. A convenient operational count is not proof of accuracy or fairness.
Illustrative case study
Fictional case: Alder Repairs uses a model to prioritize incoming maintenance reports. The model team publishes a performance report for a threshold of 0.50, but the service team raises the deployed threshold to reduce the number of review items.
During a joint review, they recognize that the published error rates no longer describe the operating pipeline. They document the deployed rule, evaluate its effects on missed urgent reports and review the capacity constraint directly. Future threshold changes receive an appropriate check and a version record.
The case demonstrates that operational changes after inference can alter important outcomes. It does not recommend a threshold or report a real-world improvement.
Common misconceptions and mistakes
- An unchanged model means an unchanged system. Output transformations can materially alter decisions.
- Post-processing can repair any underlying data problem. It may address selected effects while leaving missing information or unsuitable targets unresolved.
- Matching one fairness metric proves the whole service is fair. The metric has a defined scope and may involve tradeoffs with other concerns.
Business use
Businesses should include post-processing in the system inventory, evaluation plan and change controls. Ask whether a supplier's reported results apply to raw model scores or to the exact pipeline being deployed. A threshold, filter or re-ranking rule can be as consequential as a model update.
Choose interventions based on a clear objective and evidence about the problem. Monitor the final user-facing outputs and downstream consequences, not only model performance before transformation. Where decisions affect people, involve relevant expertise in assessing the appropriateness of the rule and its data requirements.
People also ask
Does post-processing require retraining the base model?
Not necessarily. Many methods operate on outputs while leaving the base parameters fixed.
Is calibration a kind of post-processing?
A separately fitted mapping of model scores can be, although calibration is a specific objective with its own evaluation.
Is every output filter a safety guarantee?
No. Its effectiveness and limitations depend on what it checks and how it behaves in the complete system.
Check your understanding
A model's scores stay identical, but a new threshold doubles the number of flagged cases. Has the deployed decision behavior changed?
Answer: Yes. The post-processing rule changed the mapping from scores to decisions, so the complete pipeline needs an appropriate evaluation.
Related terms
Learn next
Read classification threshold and fairness metric. Draw the steps between one model's raw output and the final action, then identify which reported evaluation results cover those steps.
More in Governance, ethics and safety
- Popularity Bias - Popularity bias is a tendency for already popular items to receive disproportionate exposure while less popular items receive little representation.
- Population Bias - Population bias is a systematic mismatch between the people represented in a dataset and the population about which conclusions or predictions are intended.
- Post-hoc explanation - A post-hoc explanation is an explanation produced by analyzing an already developed AI model or its behavior.
- Practical significance - Practical significance is the importance of an observed or estimated effect in its real-world context.
- Pre-processing algorithm - A pre-processing algorithm is a defined procedure that prepares or transforms data before a model uses it.
- Prediction bias - Prediction bias is the difference between the average of a model's predictions and the average of the corresponding observed labels in a dataset.
Sources
Google Machine Learning Glossary, post-processing: https://developers.google.com/machine-learning/glossary/responsible-ai#post-processing. Fairlearn, Postprocessing: https://fairlearn.org/main/user_guide/mitigation/postprocessing.html. Relevant returned documentation supports output adjustment, fairness constraints and threshold methods. The six-score calculation and business case are original teaching examples, not benchmark results or compliance assessments.
Previous: Post-hoc explanation Next: Post-processing algorithm