DataTerminology.comNEW
Also known as: Prediction score
A numerical value output by a model indicating the strength or confidence of a predicted outcome.
DESCRIPTION
In machine learning, a score is a continuous value that reflects a model’s confidence or affinity for a particular class or outcome before or after normalization to a probability. Scores support ranking, thresholding, and evaluation workflows. They can be logits, margins, or probabilities depending on the model and post-processing. Proper calibration aligns scores with true likelihoods, improving decision thresholds. In deployment, scoring also refers to applying a model to generate predictions on new data.
ALTERNATIVE DEFINITIONS
Google Developers Machine Learning Glossary
A value produced by a model that indicates the strength of a prediction for a particular class, which may be converted to a probability.
Highlights pre-probability nature of scores and conversion.
Microsoft Learn
Scoring is the process of using a trained machine learning model to generate predictions on new data.
Defines scoring as the operational act of prediction generation.
RELATED TERMS
Relative Frequency Confidence Score Logit Threshold Decision Function Softmax ROC Curve Precision Recall Calibration
COMPETING TERMS
RELATIONS
- Narrower: AUC Score BERTScore BLEU Score Brier Score F-beta Score F1 Score Net Promoter Score Propensity Score R2 Score ROUGE Score Silhouette Score
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Classifier (machine learning)
- Google Developers - Glossary entry defining model scores
- scikit-learn - Docs on decision_function and predict_proba
- Microsoft - Guide to model scoring in production
TOP QUESTIONS
How does a Score differ from a probability?
A score is often a raw or scaled value; probabilities are normalized to [0,1]. Scores may be converted to probabilities via softmax or calibration.
Category: Misunderstanding
Why are Scores useful for thresholding?
Scores provide a ranking that helps set thresholds balancing precision and recall for specific business trade-offs.
Category: Application
What is calibration of Scores?
Calibration aligns scores with true outcome frequencies, improving decision-making. Methods include Platt scaling and isotonic regression.
Category: Best practices
Are higher Scores always better?
Within a model’s context, higher scores indicate stronger evidence for a class. However, compare across consistent models and ensure calibration.
Category: Misunderstanding
How do Scores appear in ranking systems?
Documents or items are sorted by score to maximize relevance; learning-to-rank models optimize these scores.
Category: Example
What risks come with misinterpreting Scores?
Treating uncalibrated scores as probabilities can mislead decisions. Regular calibration and monitoring mitigate this risk.
Category: Risks
How are Scores produced in classification?
Models output logits or probabilities per class; the highest-scoring class is selected, subject to thresholding in binary tasks.
Category: Overview
What is the role of Scores in ROC and PR curves?
Varying score thresholds generates points for ROC/PR curves, summarizing trade-offs between true and false positive rates.
Category: Requirements
How to deploy scoring in production?
Expose a scoring endpoint with versioned models, validate inputs, log outputs, and monitor latency and drift.
Category: Management