Cosine Similarity Analysis

Cosine similarity analysis measures how similar two numeric vectors are by computing the cosine of the angle between them, which reflects the pattern of values regardless of their magnitude.

코사인 유사도 분석은 두 수치 벡터 사이 각도의 코사인을 계산하여 크기와 관계없이 값의 패턴이 얼마나 비슷한지를 측정하는 방법이다.

How do I measure how similar two profiles, documents or embedding vectors are regardless of their size?

두 프로필, 문서 또는 임베딩 벡터가 크기와 상관없이 얼마나 비슷한지 어떻게 측정하나요?

Cosine similarity measures the angle between two vectors, so it compares the pattern of values rather than their overall magnitude. It fits text term-frequency vectors, embeddings and user or item profiles, where length differences should not drive similarity.

When to use Cosine Similarity Analysis

• Comparing documents represented as term-frequency or TF-IDF vectors
• Comparing embedding vectors from language or recommendation models
• Finding the most similar respondents, products or items in a feature matrix

When to use something else

• Absolute differences in magnitude matter: use Euclidean distance
• You need a statistical measure of linear association between two variables: use Pearson correlation

Assumptions and requirements

• Numeric feature vectors of equal length
• Non-zero vectors
• Features on comparable scales or appropriately weighted

What the analysis reports

• Pairwise cosine similarity matrix
• Ranked list of most similar items
• Heatmap of similarities
• Cosine distance (1 − similarity) for clustering

How to report it

The mean cosine similarity between [group A] and [group B] profiles was [value] (SD = [value]).

Related methods

• Euclidean Distance
• Pearson Correlation
• Jaccard Similarity

Run Cosine Similarity Analysis on Quorant