DataTerminology.comNEW
A method that maps features into a fixed-dimensional space via hashing, enabling scalable vectorization with possible collisions.
DESCRIPTION
Feature hashing uses a hash function to transform arbitrary feature names into indices of a fixed-size sparse vector. It eliminates the need for a global dictionary, reducing memory and preprocessing cost. Collisions are expected and manageable with sufficient dimensionality or signed hashing. Popular in text and clickstream pipelines, it enables online learning and high-cardinality feature handling. Libraries like scikit-learn and Vowpal Wabbit provide efficient implementations.
ALTERNATIVE DEFINITIONS
scikit-learn
Feature hashing projects features into a low-dimensional space using a hash function, allowing scalable and memory-efficient vectorization.
Highlights scalability and memory efficiency in ML preprocessing.
Wikipedia
Feature hashing, also known as the hashing trick, is a method of vectorizing features that uses a hash function to map features to indices.
Concise definition describing mapping of features to indices.
RELATED TERMS
Count Vectorizer Term Frequency–Inverse Document Frequency Hash Function Sparse Matrix Collision Vowpal Wabbit
COMPETING TERMS
One-Hot Encoding Dictionary Vectorizer
KNOWLEDGE AREAS
AI Data Science Data Engineering MLOps
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer Data Engineer
EXTERNAL SOURCES
- Wikipedia - Feature hashing
- scikit-learn - Documentation of feature hashing
- Vowpal Wabbit - Online learning using hashing
TOP QUESTIONS
Why use the Hashing Trick?
It scales to high-cardinality features without building a dictionary, enabling fast, memory-efficient preprocessing and online learning.
Category: Relevancy
What are collisions in Feature Hashing?
Different features can map to the same index. With adequate dimensionality or signed hashing, collision impact is minimized.
Category: Risk
How do I choose the hashed vector size?
Select based on expected cardinality and acceptable collision rate; larger sizes reduce collisions at increased memory cost.
Category: Requirements
Is Feature Hashing good for interpretability?
No. Hashing obscures feature identities, complicating model explanations and feature importance analysis.
Category: Misunderstanding
How does hashing compare to one-hot encoding?
Hashing avoids dictionary construction and scales better, while one-hot preserves interpretability and exact feature identities.
Category: Comparison
Can Feature Hashing be combined with TF-IDF?
Yes. Hash tokens into indices and weight with TF-IDF to capture term importance in a fixed space.
Category: HowTo
What libraries implement Feature Hashing?
Scikit-learn’s FeatureHasher and Vowpal Wabbit provide efficient implementations, including signed hashing.
Category: Overview
How to monitor hashing impact?
Track collision rates via approximate diagnostics, evaluate downstream performance, and adjust dimensionality or hashing strategies.
Category: Observability
Is hashing suitable for categorical IDs?
Yes, especially for very large ID spaces in ads or recommender systems where dictionary maintenance is infeasible.
Category: Example