DataTerminology.comNEW
A similarity measure that computes the cosine of the angle between two non-zero vectors.
DESCRIPTION
Cosine similarity quantifies orientation rather than magnitude, making it effective when vector length is uninformative, such as TF-IDF text vectors or normalized embeddings. It ranges from −1 to 1, though many applications use the 0 to 1 range after non-negativity. It underpins nearest neighbor search, clustering, recommendation, and semantic retrieval. Angular distance is a related metric derived from the cosine.
ALTERNATIVE DEFINITIONS
Wikipedia
Cosine similarity is a measure of similarity between two non-zero vectors of an inner product space that measures the cosine of the angle between them.
Emphasizes inner product space and angular interpretation.
scikit-learn
Compute cosine similarity between samples in X and Y.
Operational definition as implemented in a standard ML library.
Stanford IR Book
Cosine similarity is the normalized dot product of two vectors, often used in the vector space model for text.
Connects cosine similarity to vector space IR and normalization.
RELATED TERMS
Dot Product Term Frequency–Inverse Document Frequency Embeddings K-Nearest Neighbors Vector Space Model Angular Distance Feature Scaling
COMPETING TERMS
Euclidean Distance Jaccard Index
KNOWLEDGE AREAS
AI Data Science Semantic & Knowledge Modeling
TARGET AUDIENCE
AI Engineer Data Scientist Data Engineer Software Developer
EXTERNAL SOURCES
- Wikipedia - Cosine similarity
- scikit-learn - User guide on cosine similarity
- Gensim - Similarity queries using cosine
- Stanford University - IR book covering vector space model
TOP QUESTIONS
When should I use Cosine Similarity?
Use it when direction matters more than magnitude, such as TF-IDF vectors or normalized embeddings in semantic search.
Category: Guidance
How does Cosine Similarity compare to Euclidean distance?
Cosine focuses on angle (orientation), while Euclidean measures absolute distance; on L2-normalized vectors, they are monotonically related.
Category: Alternatives
Can Cosine Similarity be negative?
Yes, values range from −1 to 1 for centered vectors; with non-negative features like TF-IDF, it typically ranges 0 to 1.
Category: Explanation
What are common mistakes with Cosine Similarity?
Ignoring normalization or comparing unaligned vector spaces leads to misleading similarity scores.
Category: Mistakes
How is Cosine Similarity used in retrieval?
It powers nearest neighbor search over embeddings for semantic search, recommendations, and clustering.
Category: Application
Is Cosine Similarity a metric?
Cosine similarity itself is not a distance metric; angular distance derived from cosine satisfies metric properties.
Category: Explanation
Does dimensionality reduction affect Cosine Similarity?
Yes, projections can distort angles; techniques like PCA or SVD should be evaluated for downstream similarity preservation.
Category: Risks
How to scale Cosine Similarity to large datasets?
Use approximate nearest neighbor indexes (e.g., HNSW, FAISS) optimized for angular similarity.
Category: Best-Practices
What preprocessing helps Cosine Similarity on text?
Tokenization, stopword removal, TF-IDF weighting, and L2 normalization improve discriminative power.
Category: Guidance