Keyword co-occurrence analysis rests on a simple assumption: if two keywords keep appearing in the same papers, their topics are related. Aggregate that across thousands of papers and the resulting network sketches the conceptual structure of a field. The catch is that the map is only as good as the technical decisions behind it. Below is the workflow we use in practice, from choosing the keyword field to writing the Methods paragraph.
Step 1: Choose the right keyword field
Scopus and Web of Science records carry two main keyword fields, and each has trade-offs.
| Field | Origin | Strengths | Weaknesses |
|---|---|---|---|
| Author keywords | Chosen by authors | Reflect how authors position their work; close to the field's own vocabulary. | Unstandardized, many variants, missing in some records. |
| Index keywords | Assigned by the database (Scopus) or Keywords Plus (WoS) | Broader coverage, partly controlled vocabulary. | Often generic, with non-topical terms such as "human" or "article". |
| Terms mined from titles and abstracts | Text mining | Usable when keywords are missing. | Noisy; needs heavy filtering. |
For questions about research themes, author keywords usually give the most interpretable map. Index keywords work well in biomedicine, where MeSH and Emtree vocabularies are already well standardized. Whichever field you pick, state it and explain why.
Step 2: Clean the keyword data
Cleaning takes the most time and has the largest effect on the outcome. Typical operations:
- Normalize case, spacing and hyphenation ("e-learning" versus "elearning").
- Merge singular and plural forms ("supply chain" and "supply chains").
- Merge British and American spelling ("behaviour" and "behavior").
- Merge abbreviations with full forms ("CSR" and "corporate social responsibility").
- Remove non-topical terms and the search terms themselves. If you searched for "fintech", a "fintech" node will dominate the centre and tell you nothing new.
Step 3: Build a thesaurus file
Your cleaning decisions go into a thesaurus file so the software applies them consistently. In VOSviewer this is a tab-delimited text file with "label" and "replace by" columns; in bibliometrix you pass equivalent lists through the synonyms or remove.terms arguments. A sample thesaurus-terms.txt:
label replace by
supply chains supply chain
behaviour behavior
csr corporate social responsibility
study
analysis
A row with an empty "replace by" value removes that term from the map. Keep the thesaurus with your project files and submit it as supplementary material if the journal allows; it adds a great deal of transparency.
Cleaning is also where most of the hidden judgement in a keyword map lives. Should "machine learning" and "deep learning" be merged? Usually not, because they mean different things to specialists. Should "COVID-19 pandemic" and "COVID-19" be merged? Usually yes. Make these calls with the research question in mind, write them down, and apply them consistently. When you are unsure, keep the terms separate: an over-merged map hides structure, while a slightly under-merged one simply shows two nearby nodes.
Step 4: Set the minimum occurrence threshold
The threshold controls how many nodes appear. There is no universal value; the right choice depends on the size of your dataset.
| Records | Starting threshold | Target number of nodes |
|---|---|---|
| Under 500 | 2 to 3 | 40 to 80 |
| 500 to 2,000 | 5 to 10 | 60 to 120 |
| Over 2,000 | 10 to 20 or more | 80 to 150 |
Treat these as rules of thumb, not standards. What matters is trying a few values, checking whether the cluster structure is stable, and reporting the value you settled on.
One practical way to choose is to run the map at two or three thresholds and compare the clusters side by side. If the same broad themes appear each time, your structure is robust and you can pick the most readable version.
Step 5: Clustering and resolution
VOSviewer clusters items with a modularity-based algorithm controlled by a resolution parameter (default 1.00). Higher resolution gives more, smaller clusters; lower resolution gives fewer, broader ones. A minimum cluster size setting folds very small clusters into larger ones.
Too many clusters
Clusters with only a handful of nodes, hard to name, themes split into fragments.
Too few clusters
One cluster lumps together several distinct themes, so the map stops discriminating.
About right
Often four to eight clusters, each nameable in two to four words.
Two further settings shape the picture. The normalization method, association strength by default, corrects for the fact that frequent keywords co-occur with many others simply because they are frequent; leave it at the default unless you have a methodological reason to change it. The layout parameters, attraction and repulsion, change how tightly the map is packed but do not change cluster membership. Adjust them only to improve readability and mention any change in a footnote. If two runs with the same data produce visibly different maps, check that the thesaurus, threshold and counting method were truly identical before drawing conclusions.
Step 6: Overlay by year to spot emerging topics
The overlay view colours each keyword by the average publication year of the papers that use it. In the default palette, yellow marks a more recent average year and hints at an emerging topic. Be careful: averages are pulled around by volume, so a keyword used heavily in both early and recent years can look neutral. Complement the overlay with a period-by-period analysis (split the data into two or three time windows) or with the trend topics chart in bibliometrix.
Step 7: Read and label the clusters
Labelling clusters is interpretation, and it belongs to the author's expertise. A workable routine:
- List the ten heaviest keywords in each cluster.
- Read the titles and abstracts of the most cited papers that use them.
- Propose a label that captures the shared theme rather than reusing a single keyword.
- Build a cluster summary table: label, number of keywords, leading keywords, average year, representative papers.
Reporting the map in your Methods section
The Methods paragraph for a keyword map should state the software and version, the keyword field, how terms were cleaned and where the thesaurus is available, the counting method (full or fractional), the occurrence threshold, how many keywords met it out of the total, the normalization method (association strength by default), the resolution and the minimum cluster size. Donthu et al. (2021) and Zupic and Čater (2015) offer well-structured frameworks you can follow.
A short example of how such a paragraph might read, which you should adapt to your own study rather than copy:
Keyword co-occurrence analysis was performed in VOSviewer (version 1.6.x) using author keywords. Keyword variants were merged and generic terms removed with a thesaurus file (Supplementary File S1). Fractional counting was applied. Of the distinct keywords, those occurring at least eight times were retained. Links were normalized with association strength, and clustering used a resolution of 1.2 with a minimum cluster size of five.
Every value in that paragraph can be checked by a reader who has your data and thesaurus, which is exactly the point.
A worked example (hypothetical)
Imagine a dataset of about 1,800 Scopus records on circular economy in the construction sector. This scenario is purely illustrative and the numbers are invented to show the workflow, not to describe a real study.
Field and cleaning
Author keywords are chosen. The raw list holds several thousand distinct terms; a thesaurus merges spelling variants and plural forms and removes the search terms themselves.
Threshold
A threshold of 8 occurrences leaves around 100 keywords, which keeps labels readable at journal figure size.
Clustering
At the default resolution the map shows six clusters. Raising the resolution splits one large cluster into two that the author judges to be genuinely distinct, so the higher value is kept and reported.
Overlay
The overlay suggests digital tools and material passports as recent themes, which the author then checks with a simple count of papers per period.
Notice that every decision in this example is small and defensible on its own, and that each one is written down. That combination, rather than any particular parameter value, is what makes a keyword map credible to reviewers.
Where Biblio VietSCI fits in
Our keyword co-occurrence analysis service takes care of the technical steps: cleaning, thesaurus building, threshold testing, network, overlay and density maps, and a cluster summary table, delivered with every source file. Naming the clusters, interpreting the trends and writing the Discussion remain the author's work. See pricing or submit a request on the order page.
Frequently asked questions
Should I use author keywords or index keywords?
Author keywords usually give a more interpretable thematic map. Index keywords suit biomedicine thanks to standardized vocabularies. State your choice and the reason in the Methods section.
Should I remove my search terms from the map?
Usually yes. Search terms appear in almost every record, dominate the centre of the map and add no information about thematic structure.
How many clusters is reasonable?
There is no fixed number, but four to eight clusters are typically easy to interpret. Tune the resolution and minimum cluster size, then report the settings.
Does an overlay by year prove a topic is emerging?
Not on its own. It shows the average publication year only. Combine it with a period analysis or trend chart and interpret cautiously.