Cosine Similarity Calculator
Solve cosine similarity problems step-by-step with our free calculator. See formulas, worked examples, and clear explanations.
Reviewed for accuracy by Manoj Kumar, Mathematics Educator
Cosine Similarity Calculator
Calculator
Adjust values & calculateEnter your values below. Every result is computed in your browser — no data is sent to any server.
Formula: cos(theta) = (A . B) / (|A| * |B|)
Worked example — Cosine Similarity: 0.9746 | Angle: 12.93 degrees | Cosine Distance: 0.0254
Formula
cos(theta) = (A . B) / (|A| * |B|)
Where A . B is the dot product of vectors A and B (sum of element-wise products), |A| is the magnitude of A (square root of sum of squared elements), and |B| is the magnitude of B. The result ranges from -1 to 1.
Worked Examples
Example 1: Text Document Similarity
Problem:Two document vectors are A = (1, 2, 3) and B = (4, 5, 6). Find the cosine similarity.
Solution:Dot product: 1*4 + 2*5 + 3*6 = 4 + 10 + 18 = 32 |A| = sqrt(1 + 4 + 9) = sqrt(14) = 3.7417 |B| = sqrt(16 + 25 + 36) = sqrt(77) = 8.7749 Cosine similarity = 32 / (3.7417 * 8.7749) = 32 / 32.8330 = 0.9746 Angle = arccos(0.9746) = 12.93 degrees
Result:Cosine Similarity: 0.9746 | Angle: 12.93 degrees | Cosine Distance: 0.0254
Example 2: Orthogonal Vectors
Problem:Compare vectors A = (1, 0, 0) and B = (0, 1, 0) for similarity.
Solution:Dot product: 1*0 + 0*1 + 0*0 = 0 |A| = sqrt(1 + 0 + 0) = 1 |B| = sqrt(0 + 1 + 0) = 1 Cosine similarity = 0 / (1 * 1) = 0 Angle = arccos(0) = 90 degrees
Result:Cosine Similarity: 0.0000 | Angle: 90.00 degrees | Vectors are perpendicular
Frequently Asked Questions
What is cosine similarity and how does it measure vector similarity?
Cosine similarity is a metric that measures the cosine of the angle between two non-zero vectors in a multi-dimensional space. It ranges from -1 (completely opposite directions) through 0 (perpendicular, no similarity) to 1 (identical direction, maximum similarity). Unlike Euclidean distance, cosine similarity focuses on the orientation of vectors rather than their magnitude, making it particularly useful when the length of vectors is not meaningful. For example, in text analysis, two documents about the same topic might have different word counts but similar word frequency proportions, resulting in high cosine similarity despite different vector magnitudes.
How is cosine similarity calculated mathematically?
Cosine similarity is calculated by dividing the dot product of two vectors by the product of their magnitudes. The formula is: cos(theta) = (A dot B) / (|A| * |B|). The dot product A dot B is the sum of element-wise products: a1*b1 + a2*b2 + ... + an*bn. The magnitude |A| is the square root of the sum of squared elements: sqrt(a1^2 + a2^2 + ... + an^2). This calculation works for vectors of any dimensionality, from 2D to thousands of dimensions. The result is always between -1 and 1, providing a normalized measure of directional similarity that is independent of scale.
What is the difference between cosine similarity and cosine distance?
Cosine distance is simply 1 minus the cosine similarity, converting the similarity measure into a distance metric. While cosine similarity of 1 means identical direction (most similar), cosine distance of 0 means the same thing (closest distance). Cosine distance ranges from 0 to 2, where 0 means identical direction, 1 means perpendicular, and 2 means opposite directions. Cosine distance is preferred in clustering algorithms and nearest-neighbor searches because distance metrics typically expect lower values to indicate closer items. However, cosine distance is technically a semi-metric rather than a true metric because it does not satisfy the triangle inequality in all cases.
Where is cosine similarity used in machine learning and NLP?
Cosine similarity is extensively used in natural language processing (NLP) and machine learning for comparing text documents, word embeddings, and feature vectors. In information retrieval, search engines use cosine similarity to rank documents by relevance to a query vector. Word embedding models like Word2Vec and GloVe use cosine similarity to find semantically similar words. In recommendation systems, cosine similarity identifies users with similar preference patterns. Sentence transformers and BERT models produce embeddings where cosine similarity measures semantic relatedness. It is also used in image recognition, plagiarism detection, and anomaly detection across many domains.
Why is cosine similarity preferred over Euclidean distance for text analysis?
Cosine similarity is preferred for text analysis because it is invariant to the magnitude of vectors, focusing purely on their direction. In text processing, document vectors can have very different magnitudes simply because documents have different lengths, but they may discuss the same topics with similar word frequency proportions. Euclidean distance would consider these documents dissimilar due to magnitude differences, while cosine similarity correctly identifies them as similar. Additionally, cosine similarity handles high-dimensional sparse vectors (common in bag-of-words representations) more effectively and provides a bounded output between -1 and 1, making interpretation and threshold-setting straightforward.
What does it mean when cosine similarity is zero or negative?
A cosine similarity of zero means the two vectors are perpendicular (orthogonal) to each other, indicating no linear relationship in their directions. In text analysis, this means the documents share no common terms or features. A negative cosine similarity means the vectors point in generally opposite directions, which can occur when features have meaningful negative values (like in centered or normalized data). However, in many practical applications such as term frequency vectors where all values are non-negative, cosine similarity is always between 0 and 1, and negative values cannot occur. The interpretation depends heavily on how the input vectors were constructed.
How does the number of dimensions affect cosine similarity?
The number of dimensions is the length of the vectors being compared, and it affects the computation in several ways. In very high dimensions (thousands or more), random vectors tend to become nearly orthogonal with cosine similarities close to zero, a phenomenon known as the curse of dimensionality. This means that in high-dimensional spaces, small positive cosine similarities can actually be quite meaningful. Dimensionality reduction techniques like PCA, t-SNE, or autoencoders are sometimes applied before computing cosine similarity to focus on the most informative features and reduce noise. The formula itself works identically regardless of dimensionality.
Can cosine similarity be used with sparse vectors?
Yes, cosine similarity is particularly well-suited for sparse vectors, which are vectors where most elements are zero. This is common in text processing where bag-of-words representations have thousands of dimensions but only a few non-zero entries per document. The computation can be optimized by only considering non-zero elements, since zero values do not contribute to the dot product. Sparse matrix libraries like SciPy in Python handle this efficiently, computing cosine similarity without materializing full dense vectors. This makes cosine similarity scalable to very large vocabularies and feature spaces that would be impractical with dense distance computations.
How does cosine similarity relate to the angle between vectors?
Cosine similarity is literally the cosine of the angle between two vectors, which is why it is called cosine similarity. An angle of 0 degrees means the vectors point in the same direction (cosine = 1, maximum similarity). An angle of 90 degrees means they are perpendicular (cosine = 0, no similarity). An angle of 180 degrees means they point in opposite directions (cosine = -1, maximum dissimilarity). The angle itself can be recovered using the inverse cosine (arccos) function applied to the similarity value. This geometric interpretation makes cosine similarity intuitive and visually meaningful, especially when working with 2D or 3D vector visualizations.
What are alternatives to cosine similarity for measuring vector similarity?
Several alternative similarity and distance metrics exist for different use cases. Euclidean distance measures the straight-line distance between vector endpoints and is sensitive to magnitude. Manhattan distance sums absolute differences along each dimension. Jaccard similarity measures overlap between sets and works well for binary features. Pearson correlation measures linear correlation and is equivalent to cosine similarity on centered data. Minkowski distance generalizes both Euclidean and Manhattan distances. For probabilistic distributions, KL divergence and Jensen-Shannon divergence are commonly used. The best choice depends on data characteristics: cosine for directional similarity, Euclidean for spatial proximity, and Jaccard for set overlap.
References
Reviewed for accuracy by Manoj Kumar, Mathematics Educator · Editorial policy
Related Calculators
🧮Cosine Calculator
Calculate cosine with inputs, formulas, and instant results.
🧮Annulus Area Calculator
Calculate annulus area with inputs, formulas, and instant results.
🧮Area Calculator
Calculate area with inputs, formulas, and instant results.
🧮Area of a Rectangle Calculator
Calculate the area, perimeter, and diagonal of a rectangle. Find missing sides from known area. Convert between metric and imperial area units.
🧮Area of Crescent Calculator
Calculate area of crescent with inputs, formulas, and instant results.
🧮Center of Mass Calculator
Calculate center of mass with inputs, formulas, and instant results.
🧮Centroid Calculator
Calculate centroid with inputs, formulas, and instant results.
🧮Chord Length Calculator
Calculate chord length with inputs, formulas, and instant results.