Is the overlap coefficient symmetric?

Yes. Swapping the two sets leaves both the intersection size and the smaller-set denominator unchanged.

Does a score of one prove the sets are equal?

No. It proves containment of the smaller set within the larger one for nonempty sets under this definition. Equality is only one such case.

Can it recognize synonyms automatically?

No. Ordinary set comparison uses the elements supplied. Synonym mapping or other semantic processing must happen separately and can affect the resulting score.

AI Glossary / Retrieval and knowledge systems / Beginner

Overlap coefficient

Definition

The overlap coefficient measures how much two nonempty sets share relative to the size of the smaller set. It is calculated as the number of shared elements divided by the smaller set's size. Also called the Szymkiewicz-Simpson coefficient, it ranges from zero to one and reaches one whenever either set is contained in the other.

What it means

The measure answers a containment-oriented question: how fully is the smaller collection represented in the larger one? This differs from asking how similar the collections are overall. A short list can receive a perfect overlap score against a much larger list even when the larger list contains many additional elements.

In information retrieval, the sets might contain document terms, tags or neighboring nodes in a graph. The score depends on what counts as an element and how those elements are normalized. It does not directly measure semantic similarity, relevance or the probability that two records describe the same thing.

How it works

For sets A and B, count the distinct shared elements to obtain the size of their intersection. Divide that count by min(|A|, |B|), where the vertical bars denote the number of elements in a set.

Repeated occurrences do not increase membership counts in the ordinary set-based coefficient. If a document contains the same term ten times, that term still contributes one element to its vocabulary set. Order is also ignored.

Both sets must be nonempty for the stated formula to have a nonzero denominator. An application may define a convention for empty inputs, but it should document that convention rather than imply that the formula settles it. Weighted or multiset variants are separate definitions and should not be assumed to behave identically.

Three practical examples

1. A search tool compares a short query-term set with a document's term set to identify documents containing all query terms, while using additional signals to rank them.

2. A catalog team compares the tags of two records to find cases where one record's tag set is contained within another's, then reviews possible duplication.

3. A graph analyst compares two nodes' neighbor sets to investigate shared connections, interpreting high overlap alongside the sizes of both neighborhoods.

Think of it this way

Think of a small shopping list and a large pantry inventory. If the pantry contains every item on the list, coverage of the smaller collection is complete. That does not mean the pantry and shopping list contain the same number of items, nor that everything in the pantry was requested.

Worked example

Let A = {apple, pear, plum} and B = {apple, pear, plum, peach, grape, melon}. Their intersection contains three elements, and the smaller set has size three. The overlap coefficient is therefore 3/3 = 1.

For comparison, their union contains six elements, so their Jaccard similarity is 3/6 = 0.5. The two results differ because the denominators answer different questions: full containment of the smaller set versus shared membership relative to the combined collection.

Now compare A with C = {apple, pear, kiwi, orange}. Their intersection has size two and the smaller set has size three, giving overlap 2/3, approximately 0.6667. Swapping A and C does not change the score. All calculations use distinct exact tokens; they would change if the application altered its normalization or element definition.

Illustrative case study

Fictional case study: Orchard Index uses tag sets to suggest catalog records that deserve review. A concise record with three tags and a detailed record with twelve tags receive an overlap coefficient of one because all three concise tags occur in the detailed record.

The reviewer finds that the records describe related but different products. The team therefore retains overlap as a candidate-generation signal and adds identifiers, descriptions and human review before merging records. Its review screen also displays both tag counts, making the reason for a high score easier to understand.

This fictional case illustrates a limitation of containment-based matching. It does not claim any measured accuracy or justify automatic merging from an overlap threshold alone.

Common misconceptions and mistakes

A perfect score does not require identical sets; subset containment is sufficient. Likewise, a high score on a tiny set may provide little evidence of meaningful similarity. A one-element set obtains a score of one against any nonempty set containing that element.

Do not confuse the overlap coefficient with the raw intersection count or Jaccard similarity. Avoid assuming that repeated words strengthen the ordinary set score, and handle empty inputs deliberately. Also, matching identifiers and matching natural-language tokens require different preprocessing choices.

Business use

Businesses can use overlap to identify coverage, shortlist related records and inspect shared associations. Its interpretation is especially useful when collection sizes differ and the question concerns containment of the smaller collection.

Before deployment, examine examples across a range of set sizes. Consider displaying intersection size and both input sizes alongside the normalized score. A threshold that seems selective for large tag sets may admit many uninformative matches among very small sets. Evaluate the complete workflow using reviewed cases, and avoid treating the coefficient as a substitute for permissions, eligibility rules or record identity checks.

Check your understanding

Set A has four elements, set B has ten, and they share three. What is the overlap coefficient, and does it show complete containment?

Answer: The coefficient is 3/4 = 0.75. Complete containment of the smaller set would require all four of A's elements to be shared, giving a score of one.

Related terms

Tokenization
set intersection
subset
Jaccard similarity
Dice coefficient
document similarity
graph neighbourhood and containment

Learn next

calculate overlap and Jaccard scores for identical sets, partial matches and a small subset of a large set. Then repeat the exercise after a stated normalization step, keeping a record of which element matches changed and why.

Sources

Primary sources: Neo4j Graph Data Science, Node Similarity, including the overlap coefficient: https://neo4j.com/docs/graph-data-science/current/algorithms/node-similarity/

Neo4j Graph Data Science Client, Similarity Algorithms tutorial, overlap formula and interpretation: https://neo4j.com/docs/graph-data-science-client/current/tutorials/similarity-algorithms/

Previous: Overlap - add method Next: Overlap - save method