Measure the similarity between two vectors in n-dimensional space. Widely used in machine learning, NLP, recommendation systems, and clustering to find relationships between data points.
2026-03-28T00:00:00Z
Range: -1.0 (opposite) to 1.0 (identical)
Cosine similarity measures the angular distance between two vectors in n-dimensional space, quantifying how similarly oriented they are regardless of magnitude. The metric calculates the cosine of the angle between vectors: a value of 1.0 indicates identical direction (perfect alignment), 0 means orthogonal (perpendicular, no relationship), and −1.0 means opposite directions. Unlike Euclidean distance, which emphasizes magnitude, cosine similarity focuses purely on direction, making it invaluable when comparing data with varying scales. The formula is: cosine(θ) = (A · B) / (||A|| × ||B||), where the numerator is the dot product and denominator is the product of vector magnitudes. In machine learning, cosine similarity dominates natural language processing (document/embedding comparisons), recommendation systems (user/item preference comparisons), image clustering, and anomaly detection.
Note on zero vectors: Mathematically, the cosine similarity of a zero vector (all elements = 0) with any vector is undefined, as division by zero occurs. This calculator returns 0.0000 as a practical convention—treating an empty vector as orthogonal to all others—which is often the intended behavior in applications like document similarity (empty document = no words = no relationship). However, be aware this is a convention, not a rigorous mathematical result. In production systems, zero-vector cases should be handled explicitly depending on domain requirements.
Real-world applications leverage cosine similarity’s scale-invariance: comparing two documents of different lengths but identical word ratios yields similarity 1.0, while Euclidean distance would artificially penalize length. Content-based recommendation systems calculate cosine similarity between user preference vectors and item feature vectors to suggest similar items. Collaborative filtering calculates similarity between users to find like-minded consumers for cross-recommendations. NLP embeddings (Word2Vec, BERT, GPT) use cosine similarity to find synonyms and semantic relationships: "king" and "queen" embeddings are nearly parallel despite different magnitudes. In clustering algorithms like K-means variants and hierarchical clustering, cosine distance (1 − cosine similarity) serves as the dissimilarity metric. Practical considerations include handling zero vectors (undefined similarity), choosing appropriate thresholds for "similarity" (context-dependent: >0.8 vs. >0.5), and understanding that cosine similarity assumes vector elements are meaningful (works well for counts, embeddings; less for spatial coordinates where magnitude matters).
Convert your data into numerical vector form. Examples: document text becomes word counts or TF-IDF scores (each dimension = word), user preferences become rating vectors (each dimension = item), or embeddings from neural networks. Each vector must have the same dimensionality in your comparison. For vectors—A = [1, 2, 3] and B = [4, 5, 6]—A has 3 dimensions, so B must also have exactly 3 dimensions. Missing dimensions are implicitly zero.
Multiply corresponding elements of both vectors and sum the results. For A = [1, 2, 3] and B = [4, 5, 6]: dot product = (1×4) + (2×5) + (3×6) = 4 + 10 + 18 = 32. The dot product captures how aligned the vectors are: large positive values indicate similar directions, zero indicates orthogonality, negative values indicate opposite directions. This calculation is fast even in high dimensions, making cosine similarity computationally efficient.
Compute the Euclidean norm (length) of each vector: ||A|| = √(sum of squares). For A = [1, 2, 3]: ||A|| = √(1² + 2² + 3²) = √14 ≈ 3.742. For B = [4, 5, 6]: ||B|| = √16 + 25 + 36 = √77 ≈ 8.775. Magnitudes normalize the vectors, ensuring the cosine similarity depends only on direction, not scale. A vector scaled 2× in all dimensions has the same cosine similarity to all other vectors—only the angle matters.
Divide the dot product by the product of magnitudes: similarity = (A · B) / (||A|| × ||B||). Using our example: similarity = 32 / (3.742 × 8.775) ≈ 32 / 32.85 ≈ 0.9742. The result is a single number ranging from −1 to 1. Check for edge cases: if either magnitude is zero (null vector), the similarity is undefined; most implementations return 0 in this case. Result clamping ensures output stays in the valid range.
Similarity >0.8 typically indicates strong relationship (context-dependent). Common thresholds: 0.9+ = nearly identical, 0.7–0.9 = highly similar, 0.4–0.7 = moderately similar, 0–0.4 = weakly related, <0 = opposing directions. In recommendation systems, find top-K similar items/users by calculating all pairwise similarities and sorting. Document search: find articles similar to a query by computing query vector similarity against document corpus. Fine-tune thresholds based on domain: stricter (0.9+) for precision-critical tasks, looser (0.6+) for recall-heavy applications.
Recommendation system: find products similar to a user’s favorite based on feature vectors:
Each dimension normalized to 0–100 scale for fair comparison
Euclidean distance is sensitive to vector magnitude. Two documents with identical word ratios but different lengths have high Euclidean distance. Cosine similarity focuses on direction, making it ideal when magnitude is irrelevant (documents, embeddings, preference vectors). Choose based on your use case: Euclidean when scale matters, cosine when only alignment matters.
Vectors are orthogonal (perpendicular, 90° angle), indicating no correlation or overlap. In document analysis, zero similarity means no common terms. In recommendation systems, users with zero similarity share no preference overlap. It’s neither good nor bad—just indicates independence. Negative similarity indicates opposite directions.
Yes, when vectors point in opposite directions (>90° angle). Example: [1, 1] vs. [−1, −1] has similarity −1. However, many applications (document similarity, ratings) use non-negative vectors, so results range 0–1. Negative similarity is common in statistical applications where dimensions can have opposing signs.
Convert text documents to vectors using TF-IDF or embeddings (Word2Vec, BERT). Calculate similarity between query vector and each document vector. Rank results by similarity score. Higher scores surface more relevant documents. Embeddings capture semantic meaning, so similar concepts have high cosine similarity even with different wording.
Cosine distance = 1 − cosine similarity. Both measure separation; distance uses subtraction, similarity uses raw scores. In clustering algorithms, use cosine distance (0 for identical, 2 for opposite). Many clustering libraries support both; choose based on algorithm requirements. Higher distance = more dissimilar.
Zero vectors (all zeros) have undefined similarity—the formula has zero in the denominator. Most implementations return 0, treating null vectors as maximally dissimilar. Filter zero-magnitude vectors before calculations. In machine learning, zero vectors often indicate missing data; handle appropriately (remove, impute, or flag).
Depends on domain. Document similarity: >0.8 very similar, 0.5–0.8 similar, <0.3 dissimilar. Recommendations: >0.7 strong match. Information retrieval: >0.6 relevant. Start conservative (higher threshold), lower if too few matches, then observe precision/recall trade-off. Thresholds vary by data and application.
Yes—cosine similarity works in any dimension. Computational complexity is O(n) per pair comparison (n = dimensionality). With modern hardware, thousands of dimensions are fast. However, high-dimensional spaces exhibit “curse of dimensionality”: most vectors become nearly orthogonal, reducing semantic meaning. Use dimensionality reduction techniques (PCA, embeddings) first.
Related Tools
Generate bold Unicode text.
Calculate paper thickness.
Calculate paper weight.
Calculate parallel speedup.
Calculate growth doubling time.
Calculate website bounce rate.