Bibliometric Visualisation with VOSviewer

Discover patterns, drive research excellence

VOSviewer will draw a map from almost any export you give it. Whether that map means anything depends on decisions made before the software is opened, and on knowing which of its four views answers which question.

The five outputs on the poster are the ones worth producing: network visualisation, overlay visualisation, density visualisation, cluster analysis, and trend and thematic evolution. They are not alternative styles of the same picture. Each answers a different question, and using the wrong one is how a review ends up with a colourful figure that supports no argument.

The network view shows structure: which terms or authors occur together, and how the field divides into communities. The overlay view colours the same network by average publication year, which turns a static map into a history — the blue regions are where the field has been, the yellow regions are where it is going. For identifying an emerging gap, the overlay is by far the most useful of the four.

The density view shows concentration rather than structure, and is best for a single summary figure in a paper where space is tight. Cluster analysis is the numerical backbone underneath all of them, and its resolution parameter — how finely the field is divided — is a choice you must report, because a different value produces a different set of clusters.

The thesaurus file is what separates a good map from a bad one

Raw keyword data is messy in entirely predictable ways. 'Machine learning', 'machine-learning' and 'ML' appear as three separate nodes. British and American spellings split a term in half. Plurals fragment counts. An author appears under three variants of their own name. Every one of these makes the map wrong, and none of them is visible unless you look.

The fix is a thesaurus file — a simple two-column list telling VOSviewer which labels to merge. Building it takes an hour or two of careful reading and it transforms the result: clusters sharpen, spurious small nodes disappear, and the counts start meaning what they claim to mean. We build it, and we hand it over with the project so the analysis can be rerun.

The same discipline applies to thresholds. A minimum occurrence of five will produce a different map from a minimum of ten, and both may be defensible — but the value must be stated. A reviewer cannot reproduce your figure without it, and increasingly they will ask.

Writing about a map without over-reading it

A network map is a picture of co-occurrence, not of causation and not of importance. Two terms sitting close together means they appear in the same documents; it does not mean one causes the other, and it does not mean the relationship is well studied. Node size reflects frequency, which reflects how often people write about something — not how well it is understood.

A sparse region between two dense clusters is the most interesting feature a map produces, and also the easiest to over-claim. It may be a genuine gap. It may be an artefact of the search string. It may be a field that exists under different terminology your search did not capture. The honest treatment is to name the sparse region, offer the gap interpretation, and then check it with a targeted search before building an argument on it.

We write the interpretation to that standard, and we say in the limitations what the mapping method cannot see. Reviewers of bibliometric papers are increasingly alert to over-claiming, and a paper that draws its own boundaries reads as far more competent than one that does not.