DataTerminology.comNEW
A calibration method fitting a logistic regression on model scores to produce calibrated probabilities.
DESCRIPTION
Platt scaling calibrates classifier scores by fitting a sigmoid (logistic regression) on validation data. It is commonly applied to SVM decision values but generalizes to other models. The approach is parametric, simple, and data-efficient. It may underfit when calibration requires complex, non-monotonic mappings. Proper use requires a separate validation set to avoid optimistic bias. Calibration quality is evaluated with reliability diagrams and Brier score.
ALTERNATIVE DEFINITIONS
Wikipedia
Platt scaling is a method for transforming the outputs of a binary classifier into calibrated probabilities by fitting a logistic regression model to the outputs.
Explicitly mentions binary classifiers and logistic regression mapping.
scikit-learn
Platt scaling consists of fitting a logistic regression model to the scores of a classifier.
Concise API-oriented description of the technique.
RELATED TERMS
Calibration Logistic Regression Calibration Curve Brier Score CalibratedClassifierCV Support Vector Machine Cross-Validation Model Overfitting
COMPETING TERMS
Isotonic Regression Temperature Scaling
RELATIONS
- Broader: Scaling
KNOWLEDGE AREAS
AI Data Science MLOps Testing & Validation
TARGET AUDIENCE
Data Scientist MLOps Engineer AI Engineer
EXTERNAL SOURCES
- Wikipedia - Platt scaling
- scikit-learn - Calibration methods overview
- Papers With Code - Method summary and papers
TOP QUESTIONS
When should I use Platt Scaling?
Use it to calibrate scores from models like SVMs or boosted trees when probabilities are miscalibrated but rankings are reasonable.
Category: Guidance
How does Platt Scaling compare to Isotonic Regression?
Platt is parametric and data-efficient; isotonic is nonparametric and more flexible but prone to overfitting with small data.
Category: Comparison
What data is needed for Platt Scaling?
A separate validation set with true labels and model scores to fit the logistic mapping without leakage.
Category: Requirements
Does Platt Scaling work for multiclass problems?
Apply one-vs-rest calibration per class or use multiclass calibration schemes; ensure probabilities sum to one.
Category: Advanced
How do I evaluate calibration after Platt Scaling?
Use reliability diagrams, expected calibration error, and Brier score to assess improvements.
Category: Evaluation
What’s a common mistake with Platt Scaling?
Fitting the sigmoid on training data, which inflates performance due to leakage.
Category: Mistakes
Can I combine Platt Scaling with cross-validation?
Yes, fit calibration on out-of-fold predictions to avoid bias and use all data efficiently.
Category: Best practices
How sensitive is Platt Scaling to class imbalance?
It can be affected; consider class weighting or balanced sampling for the calibration set.
Category: Risks
Is Platt Scaling always monotonic?
Yes, the logistic function is monotonic, preserving score orderings.
Category: Concept