DataTerminology.comNEW
Also known as: Count Encoding
Encoding each category with its occurrence frequency or count within the dataset.
DESCRIPTION
Frequency encoding replaces categorical levels with their counts or relative frequencies, reducing dimensionality while reflecting category prevalence. It is simple and effective for high-cardinality features and models that handle numeric inputs. However, it can correlate with the target when computed improperly, causing leakage. It may overweight popular categories without capturing nuanced relationships. Use careful cross-validation and compute frequencies strictly on training folds.
ALTERNATIVE DEFINITIONS
category_encoders
CountEncoder replaces categories with the number of times they appear.
Emphasizes replacement by category occurrence count.
Kaggle Learn
Count encoding replaces a categorical value with the number of times it appears in the dataset.
Introduces frequency-based encoding in an applied learning context.
RELATED TERMS
Label Encoding Hashing Trick Categorical Data CatBoost Encoding
COMPETING TERMS
One-Hot Encoding Target Encoding
RELATIONS
- Broader: Encoding
KNOWLEDGE AREAS
Data Science AI MLOps Data Engineering
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Data Engineer Analytics Engineer
EXTERNAL SOURCES
- Wikipedia - Categorical variable
- category_encoders - Count/frequency encoder documentation
- Kaggle Learn - Course covering categorical encodings
- scikit-learn - General categorical encoding overview
TOP QUESTIONS
When should Frequency Encoding be used?
Use it for high-cardinality categorical features to reduce dimensionality while preserving category prevalence. It often works well with tree-based models.
Category: Requirements
What is a common mistake with Frequency Encoding?
Computing frequencies on the full dataset causes leakage. Always compute on training folds and apply to validation and test sets.
Category: Mistakes
How does Frequency Encoding compare to One-Hot Encoding?
Frequency encoding produces one numeric column per feature; one-hot creates many binary columns. One-hot avoids prevalence bias but increases dimensionality.
Category: Alternatives
Is Frequency Encoding appropriate for linear models?
It can be, but prevalence may not reflect linear relationships. Regularization and interaction terms may help mitigate biases.
Category: Opinion
How do you prevent drift in Frequency Encoding?
Monitor category distributions and retrain encoders periodically. Use smoothing or min count thresholds to handle rare categories.
Category: Management
What risks are associated with Frequency Encoding?
Popular categories can dominate model behavior, and unseen categories may be misrepresented without proper handling logic.
Category: Risks
How to implement Frequency Encoding in practice?
Compute counts per category in training data, map them to categories, and optionally normalize to relative frequencies.
Category: HowTo
Does Frequency Encoding capture target relationships?
No, it captures prevalence only. If target correlation is needed, consider target encoding with careful cross-validation.
Category: Misunderstanding
What hot topic influences Frequency Encoding?
Learned embeddings and target encodings often outperform simple frequency-based approaches for complex categorical patterns.
Category: Topic