Definition
Re-scoring means adjusting or recomputing model scores (predictions, probabilities, or rankings) after the initial model output, usually by applying additional information, calibration, or rules.
- Common in ranking systems, recommendations, search engines, fraud detection, and NLP pipelines.
- Purpose: improve relevance, fairness, or calibration without retraining the whole model.
Where It’s Used
- Search & Recommendation Systems
- Base model produces initial ranking scores.
- Re-scoring applies business rules, diversity constraints, or personalization.
- Example: re-score results to boost new items, demote duplicates, or enforce diversity.
- Classification / Probabilities
- Model outputs raw probabilities.
- Re-scoring adjusts them with calibration (Platt scaling, isotonic regression, Bayesian correction).
- Ensembles
- Combine multiple models by re-scoring their outputs.
- Example: weighted average of fraud scores from different models.
- Fairness / Guardrails
- Re-scoring can apply fairness constraints.
- Example: ensure no demographic group gets systematically lower loan approval scores.
- NLP (Reranking)
- In speech recognition or machine translation, multiple candidate outputs (N-best list) are generated.
- Re-scoring chooses the best candidate using a second model (like a language model).
Examples
- Fraud Detection
- Base model score = 0.7 fraud probability.
- Re-score using customer risk profile → final fraud score = 0.85.
- E-commerce Search
- Initial ranking: relevance score from ML model.
- Re-score: add +0.2 boost if product is on promotion.
- Speech Recognition
- N-best hypotheses generated by acoustic model.
- Re-scored with language model → most fluent sentence selected.
Benefits
- Improves accuracy without full retraining.
- Adds flexibility (business rules, fairness).
- Useful for real-time adjustments (context-aware scoring).
Challenges
- Can add latency if re-scoring pipeline is complex.
- Needs careful weighting to avoid bias.
- Overuse may “patch” model weaknesses instead of solving root issues.
Summary
Re-scoring = adjusting or recomputing model outputs after initial scoring, often with extra features, calibration, or business rules.
- Common in ranking, recommendations, fraud detection, and NLP.
- Helps improve relevance, fairness, or calibration without retraining.
