Does every edge have to be weighted?

No. A simple network may record only whether the specified relationship exists, though weighted edges preserve more information about the counting rule.

Can a network prove causation?

No. Joint appearance alone does not identify a causal mechanism.

Why do two tools produce different maps?

They may use different extraction, counting, normalization, thresholds, layouts, or clustering settings even when starting from similar material.

AI Glossary / Data science and analytics / Intermediate

Co-occurrence network

Definition

A co-occurrence network represents items as nodes and connects items that appear together within a defined context. The items might be words, document keywords, products, or observed features. An edge records the specified co-occurrence relationship, and its weight may summarize how often or how strongly the pair occurs together. The meaning of the network depends on the context and counting rules used to construct it.

What it means

A pair of terms appearing in the same document is different from a pair appearing in the same sentence or within a short word window. All can produce co-occurrence networks, but they answer different questions. A link indicates the chosen form of joint appearance; it does not necessarily imply similarity, agreement, direct interaction, or causation. A diagram becomes informative only when readers know what its nodes and edges mean.

How it works

Define the collection, items, and co-occurrence unit. Clean or standardize item names where justified, preserving meaningful distinctions. Count joint appearances using a specified rule, such as once per document for each pair. Optionally normalize weights to account for item frequency or apply a justified threshold to simplify the network. Layout and clustering methods can help explore the resulting graph. Record these choices because alternative units, counting methods, and filters can produce noticeably different views of the same source material.

Three practical examples

Research discovery: a keyword network helps analysts locate themes that appear together across a defined collection of papers.
Customer feedback: a team explores which service issues are mentioned together in support cases, then reads the underlying cases to interpret the pattern.
AI project planning: a glossary team examines which concepts appear together in learning materials to suggest useful cross-links, subject to editorial review.

Think of it this way

Think of drawing lines between ingredients used in the same recipes. Frequently linked ingredients may suggest familiar combinations, but the lines do not prove that one ingredient causes the other or that every combination tastes good. The analogy explains shared context. A text network also depends on language interpretation, document boundaries, and how repeated mentions are counted.

Worked example

Consider three fictional documents. Document 1 contains the terms privacy, consent, and data. Document 2 contains privacy and data. Document 3 contains consent and audit. Count each pair at most once per document, regardless of repeated mentions within that document.
The privacy - data edge has weight 2 because the pair appears in Documents 1 and 2. The privacy - consent, consent - data, and consent - audit edges each have weight 1. There is no privacy - audit edge under this rule. If the team displays only edges with weight at least 2, only privacy - data remains visible. That does not mean the other relationships never occurred; the display threshold removed them. Nor does it prove that privacy is more important than consent. The example separates the observed counts from the choices made to display and interpret them.

Illustrative case study

In a fictional customer-service organization, analysts build a network from ticket text. They notice a strong link between delivery and refund and assume late deliveries are causing refunds. Reading a sample of tickets reveals several different stories, including customers asking about refund policies before placing orders. The team refines the text categories and uses the network to identify cases for investigation rather than to assert a causal explanation. It also documents whether automated extraction treats negated statements correctly. The network supports exploration, but the original evidence remains necessary.

Common misconceptions and mistakes

Co-occurrence is not the same as semantic similarity: words can appear together because they contrast. A common term may connect to many others simply because it appears almost everywhere. Node positions and colors often result from layout or clustering algorithms, not direct measurements of physical distance or fixed natural categories. Removing an edge through thresholding does not erase the underlying observation. Finally, automatic term extraction can merge distinct meanings or split synonyms, changing the apparent structure before visualization even begins.

Business use

Businesses can use these networks to organize a large collection and identify promising relationships for closer inspection. Start with a practical question and retain links back to the contributing records. Explain counting, filtering, and normalization in the accompanying description. For AI-assisted analysis, test extracted terms and relationships on actual source examples rather than trusting a visually convincing map. Changes between two networks may reflect different collection sizes or processing settings, so comparison requires consistent definitions and suitable normalization.

Check your understanding

An edge disappears after increasing the minimum display weight from one to two. Does that prove the pair never co-occurred?

Answer: No. It shows only that the edge failed the new display criterion. The underlying count and filtering rule must be consulted.

Related terms

Graph
Node
Text mining
Normalization
edge
term extraction
association analysis
network visualization
bibliometrics

Learn next

construct a small network manually from three or four records using a stated counting rule. Change the context from document to sentence and compare the resulting edges. Then inspect how a threshold affects the display without changing the original evidence.

Sources

Sources: The VOSviewer developers describe support for term co-occurrence networks and counting approaches: https://www.vosviewer.com/features . Their example collection includes a term co-occurrence network: https://app.vosviewer.com/docs/examples/ . Historical manual excerpts provide counting-method context: https://www.vosviewer.com/documentation/Manual_VOSviewer_1.6.8.pdf . These primary project sources were consulted for network construction concepts, not a review of current software controls. The document counts and customer-service case are original illustrations.

Previous: Co-occurrence matrix Next: Co-training