⚙️ Re-scoring

Re-scoring#

Recomputing model scores on data, e.g. after retraining or recalibration.

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

What it is#

Re-scoring is adjusting or recomputing a model’s scores — predictions, probabilities or rankings — after the initial output, by layering on extra information, calibration or rules. It improves relevance, fairness or calibration without retraining, and shows up in ranking, recommendations, search, fraud detection and NLP pipelines.

Where it’s used#

Five common settings. In search and recommendation, a base ranking is re-scored with business rules, diversity constraints or personalisation (boost new items, demote duplicates). In classification, raw probabilities are re-scored by calibration (Platt scaling, isotonic regression, Bayesian correction). In ensembles, several models’ outputs are combined — a weighted average of fraud scores, say. For fairness, re-scoring enforces guardrails so no group is systematically disadvantaged. And in NLP reranking, an N-best list from an acoustic or translation model is re-scored by a second model (a language model) to pick the most fluent candidate.

Examples#

A fraud model’s 0.7 probability is re-scored against the customer’s risk profile to 0.85; an e-commerce relevance score gets a +0.2 promotion boost; a speech system’s N-best hypotheses are re-scored by a language model to choose the most fluent sentence.

Benefits and challenges#

It buys accuracy without full retraining, flexibility (rules, fairness) and real-time, context-aware adjustment. The risks: a complex re-scoring stage adds latency, weighting must be tuned to avoid introducing bias, and leaning on it too hard can paper over model weaknesses instead of fixing their root cause.

Theme: MLOps, Serving & Monitoring · All terminology

Hint

Mind map — connected ideas

Platt Scaling · Isotonic Regression · Recalibration · Ranking Algorithms · NDCG (Normalized Discounted Cumulative Gain) · Demographic Parity (Statistical Parity)

Hint

More in MLOps, Serving & Monitoring

AWS SageMaker Endpoints · Caching · Cloud Inference · Cloud Inference with Big Payloads · Compute budgets · Continuous Retraining · Feature Values · Guardrails (in ML & Data Systems) · Inference Cost (Inference $) · Latency Guardrails · Manual review minutes · Model KPIs (Key Performance Indicators) · Model Stability · Monitoring Pipelines