AI Glossary / Data science and analytics / Intermediate
Cosine similarity
Definition
Cosine similarity measures the alignment of two nonzero numerical vectors by taking the cosine of the angle between them. It is calculated as their dot product divided by the product of their lengths. The result emphasizes direction rather than overall magnitude. In AI applications, it is often used to compare vector representations of text, images, or other items.
What it means
Two vectors can have very different lengths while pointing in the same direction, producing a cosine similarity of one. For real-valued nonzero vectors, the score ranges from minus one to one: opposite directions give minus one and perpendicular directions give zero. When all components are nonnegative, the score cannot be negative. The mathematical score becomes a useful notion of similarity only when the representation and application make that interpretation appropriate.
How it works
Represent both items in the same coordinate system with corresponding dimensions. Multiply matching components and add the products to obtain the dot product. Calculate each vector's Euclidean length, then divide by the product of those lengths. If both vectors have already been normalized to unit length, their dot product equals their cosine similarity. A zero vector has no defined direction, so the mathematical formula is undefined for it; software may apply a documented convention that requires explicit handling.
Three practical examples
- Document retrieval: a search system ranks document vectors against a query vector produced by a compatible representation method.
- Duplicate review: a team uses vector similarity to suggest potentially related records for inspection, while checking whether important details differ.
- Recommendation: a system compares item profiles, provided the chosen features and similarity measure have been evaluated for the recommendation task.
Think of it this way
Think of two arrows drawn from the same starting point. Cosine similarity asks how closely their directions align, not how far they extend. This is a precise geometric analogy for the formula. The leap from arrow direction to meaning, relevance, or user preference depends on how the original items were converted into vectors.
Worked example
Take the fictional two-dimensional vectors A = (3, 4) and B = (6, 8). Their dot product is 3 × 6 + 4 × 8 = 50. Their lengths are 5 and 10, so their cosine similarity is 50/(5 × 10) = 1. B is twice A, but the direction is unchanged.
Now compare A with C = (4, −3). The dot product is 3 × 4 + 4 × (−3) = 0, giving cosine similarity zero. This demonstrates perpendicularity in the chosen coordinates; it does not establish that two real documents represented by those vectors share no meaning. The interpretation depends on the features. If one vector is (0, 0), division by its zero length is undefined, so a system needs an intentional policy for that case.
Illustrative case study
In a fictional knowledge-base project, an AI search system retrieves two passages with high cosine similarity to a question. One passage describes the current procedure and the other describes a retired procedure using nearly identical language. Reviewers add version and validity filters and evaluate retrieval against realistic questions. They retain similarity as one retrieval signal while checking whether the selected passage is appropriate evidence. The example shows that a high score can identify related language without establishing that the content is current or supports the answer.
Common misconceptions and mistakes
A score of 0.8 is not an 80% probability that two items mean the same thing. There is no universal threshold that works across every embedding model, dataset, and task. Cosine similarity is not ordinary Pearson correlation unless additional centering conditions are applied. Normalization deliberately removes overall vector magnitude, which may discard useful information in some applications. Finally, vectors from unrelated embedding spaces cannot be meaningfully compared just because they happen to have the same number of dimensions.
Business use
Businesses should assess retrieval or recommendation quality using relevant examples and outcomes rather than treating a similarity score as self-explanatory. Document the representation model, preprocessing, and scoring rule. For AI-assisted answers, retain checks for source authority, relevance, time validity, and support for the generated claim. Model changes can alter score distributions and rankings, so existing thresholds need reassessment. A useful system connects the mathematical comparison to the user's task and makes failure cases visible during evaluation.
People also ask
Does cosine similarity ignore all differences?
No. It ignores positive overall scaling of an individual vector but remains sensitive to changes in direction and relative component values.
Can it be negative?
Yes, for vectors with suitable positive and negative components; nonnegative feature vectors produce nonnegative scores.
Is the highest-scoring result always correct?
No. It is only the highest under that representation and scoring process, and may still be irrelevant, outdated, or misleading.
Check your understanding
A vector is multiplied by five without changing the signs or proportions of its components. Does its cosine similarity with the original nonzero vector change?
Answer: No. Positive scaling preserves direction, so the similarity remains one. Other measures that use magnitude may behave differently.
Related terms
- Vector
- Embedding
- Dot product
- Normalization
- Euclidean distance
- Information retrieval
- Nearest neighbor search
- semantic search
Learn next
calculate cosine similarity for several small vectors and compare it with Euclidean distance. Then inspect retrieved text pairs with similar scores but different practical usefulness. This connects the geometry to the separate question of whether the representation serves the intended task.
More in Data science and analytics
- Continuous Uniform Distribution - A continuous uniform distribution is a probability model with constant density across a finite interval and zero density outside it.
- Continuous variable - A continuous variable is a numerical variable modeled as capable of taking any value within an interval, rather than only separate countable values.
- Contrast set learning - Contrast set learning searches for combinations of attribute values whose prevalence differs meaningfully across predefined groups.
- Cross-industry standard process for data mining - The Cross-Industry Standard Process for Data Mining, usually abbreviated CRISP-DM, is a process model for organizing data-mining projects.
- Dark data - Dark data is information an organization has collected or retained but does not meaningfully use for analytics or decision-making.
- Data cleaning - Data cleaning is the process of identifying and addressing errors, inconsistencies, and other quality problems in data so that the data are more suitable...
Sources
Sources: scikit-learn's primary documentation states the normalized dot-product formula and unit-length equivalence: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.cosine_similarity.html . Steck, Ekanadham, and Kallus examine limitations of interpreting embedding cosine scores: https://arxiv.org/abs/2403.05440 . The function documentation and research abstract were read. The abstract's model-specific findings are not generalized into a claim that cosine similarity is always invalid; the numerical vectors and business case are original examples.
Previous: Coscheduling Next: Cost efficiency