Documentation: Research and Patent Clusters Dataset

Overview

What is this dataset?

The Research and Patent Clusters Dataset (“Clusters Dataset”) contains detailed information on CSET’s research clusters and patent clusters. Our research clusters are groupings of global scientific publications from Merged Academic Corpus organized into approximately 91,000 clusters based on their citations and title and abstract text similarity. Our patent clusters are groupings of global patent documents organized into approximately 116,000 clusters based on their citations, title and abstract text similarity, and Cooperative Patent Classification (CPC) text similarity. The dataset includes publications and patents in each cluster along with cluster metadata such as titles and summaries, common subjects, top producing entities.

How do I get it

The dataset is available on Zenodo.

Which ETO products use it?

Our research and patent clusters are displayed in our Map of Science and forthcoming Map of Patents.

What are its main limitations?

  • Cluster disciplines, fields, subfields, and concepts may be imprecise. We rely on algorithms and numerical thresholds to identify relevant subjects (including disciplines and fields and emerging technology subjects) and concepts for each cluster. In practice, we find that their results generally give a good sense of subjects and concepts relevant to different clusters. However, they may not capture every relevant subject or concept, and some of the subjects and concepts they do identify may be questionable. Use caution, especially when analyzing clusters individually rather than in the aggregate.
  • Hallucinations are a known issue with LLMs, which we use to generate cluster titles and summaries and assist with publication emerging technology subject assignment. We mitigate their effect through a second-pass verifier, and check for hallucinations manually in a subset of clusters, but it is possible hallucinations could exist in the larger dataset.
  • The clusters are derived from other data and reflect the limitations of that data. Known limitations in the Merged Academic Corpus and its underlying data sources, including incomplete data for the most recent years, will exist in the cluster-level data.
  • The dataset isn’t suitable for all types of analysis. We continuously add new articles to the clusters based on their connections, but we don’t continuously add clusters, delete clusters, or change the cluster assignment of already assigned articles. This means clusters are a snapshot of the current research and patent landscape, and cannot be tracked historically in the Maps. This also means the publications within a cluster can change when new clusters are generated, which can lead to changes in cluster metadata.

What are the terms of use?

This dataset is subject to ETO's general terms of use. If you use it, please cite us.

Does it contain sensitive information, such as personally identifiable information?

No.

How do I cite it?

Please cite the "Emerging Technology Observatory Research and Patent Cluster dataset," including the link.

If you use Map of Science to access cluster data, you should cite the Map of Science instead.

Structure and content

The Clusters Dataset contains detailed information on CSET’s research clusters and patent clusters.

For the research clusters, each cluster has the following metadata fields. Note that this dataset does not include all the cluster metadata displayed on the Map of Science cluster detail pages.

  • cluster_id: a unique cluster identifier
  • cluster_title: LLM-generated cluster title
  • cluster_summary: LLM-generated cluster summary
  • n_articles: number of articles in the cluster
  • n_recent_articles: number of articles in the cluster published in the last 5 years
  • domains: top, high level domains of the cluster
  • fields: more granular fields in the cluster
  • keywords: extracted phrases based on the titles and abstracts of articles in the cluster
  • top_papers: top papers in the cluster, including titles, DOI, OpenAlex ID, year of publication, number of within-cluster citations, a core measure for that paper in that cluster, a citation ranking, a ranking for core articles, a ranking for review articles, and flags for whether or not the article is a top cited, core article, or review article,
  • author_info: most prolific authors in that cluster
  • author_orgs: top author affiliations within the cluster
  • author_country_info: country information from author organizations
  • source_info: top sources for articles in that cluster

For the patent clusters, each cluster has the following metadata fields.

  • cluster_id: a unique cluster identifier
  • cluster_title: LLM-generated cluster title
  • cluster_summary: LLM-generated cluster summary
  • n_patent_families: number of patent families in that cluster
  • main_category: main category of patents in the cluster
  • top_categories: top categories of patents in the cluster
  • keywords: extracted keywords from the patent titles and abstracts
  • top_patents: top patent families within the cluster, including family ID, a core statistic, number of within-cluster citations, a core ranking, a citation ranking, and the patent family title
  • top_inventors: inventors with the most patent families in a cluster
  • top_assignees: assignee (owners) with the most patent families in a cluster

Sources and methodology

Data sources

Our research clusters are generated from the Merged Academic Corpus, which includes over 280 million scientific publications from The Lens, OpenAlex, Semantic Scholar, and ArXiv. Our patent clusters are generated from patent documents from The Lens and PATSTAT.

Clustering methodology

A research cluster is a group of publications that are highly interconnected through direct citation and similarity in the text of their titles and abstracts. A patent cluster is a group of patent documents that are highly interconnected through direct citation, similarity in the text of their titles and abstracts, and similarity in their Cooperative Patent Classification.

We generate research clusters by first generating a large network connecting all scientific publications in our Merged Academic Corpus using the similarity between articles determined by the similarity in their embeddings using a pre-trained, multilingual SentenceTransformer model. We combine this with a network based on between-article citations, to generate a hybrid network that includes both the relevant citation information and text similarities between articles.

We generate patent clusters in the same way, but with patent families as the basic unit. Because the same invention is often filed in multiple countries and at multiple points in time, we group patents into families using PATSTAT identifiers and treat each family as a single document (excluding design and plant patents, as well as families with no title or abstract text in any language). The patent network also combines three signals rather than two: family-to-family citations, similarity of title and abstract embeddings, and similarity of embeddings of the text descriptions of each family's CPC and IPC classification codes.

We then organize documents into clusters using the Leiden algorithm. Essentially, this algorithm generates many random clusterings then finds the one whose clusters have the highest modularity.

However, because the clustering algorithm involves a degree of randomness, relying on the algorithm alone to generate the clusters can produce unstable results - that is, different runs of the algorithm will often assign the same document to different clusters. For research clusters, we address this by screening out documents that are less likely to convey meaningful information, which tend to "throw off" the clustering algorithm. Specifically, we screen out articles unless:

  • they cite at least 10 other articles and were published more than 4 years ago, OR
  • they cite at least one other article AND are cited by at least one article, AND were published more than 1,000 days ago.

We also disregard "unstable" article links when clustering articles - that is, network connections that cause the clustering structure to vary significantly over different runs of the clustering algorithm.

For patent clusters, we do not screen out individual documents; instead, we address instability through the network connections alone, running the clustering algorithm five times and keeping only the stable links - those that place the same two families in the same cluster in at least two of the five runs.

After screening documents and connections, we run the Leiden algorithm again using only the stable documents and network links that remain. After that, we delete clusters with fewer than 50 documents, and assign each of the documents in those clusters to the remaining cluster that is most strongly connected to that document. (Documents published between clustering updates are assigned to clusters using a similar method.)

Finally, each document that was screened out is added to the cluster to which it has the strongest connection. For patent clusters, this reattachment proceeds in successive waves, so that families whose only connections are to other reattached families can still be assigned.

This process of data screening, algorithmic sorting, and selective reassignment produces clusters that are (1) highly modular; (2) large enough to be analytically meaningful - generally in the hundreds of documents; and (3) stable across different runs.

Cluster metadata: creating information about clusters

For each cluster, we include metadata to describe the clusters and their constituent documents.

Cluster ID

  • Each cluster has a unique numerical ID. ID numbers are arbitrarily selected and are purely for identification.

Cluster Title and Summary

  • To generate cluster titles and summaries, we feed cluster's primary fields/categories, extracted keywords, and the titles of the top 10 core documents into Gemini 3.1 Flash-Lite Preview. The output is then run through a second-pass verifier prompt that rewords the summary to enforce neutral framing and length constraints.

Research disciplines and fields

  • Each research cluster is associated with three high-level disciplines (such as medicine or computer science) and three more specific fields (such as cardiology or soil science), determined using the article-level disciplines and fields in the cluster and their relevance scores.
    • To learn how disciplines and fields are assigned at the article level, refer to the MAC documentation.
    • To use article-level disciplines and fields to determine cluster disciplines and fields, we calculate a cluster-level relevance score, defined as the number of articles with that discipline/field in its top-three disciplines/fields, based on relevance scores for each article in the cluster. We then select the three top-scoring disciplines and three top-scoring fields as the overall disciplines and fields for that cluster. In each of these steps, we use articles from the past 10 years only. Because subfields and topics are defined hierarchically, meaning that a subfield and topic are nested under a particular field, we then consider each subfield and topic assigned to at least one article in the cluster, provided its parent field is determined to be in that cluster’s top-scoring fields. We again calculate a cluster-level relevance score, defined as the number of articles with that subject (subfield or topic) with a significant relevance score in the cluster. We then select the three top-scoring subfields and topics. However, because a small number of articles in a cluster may be connected to a particular subfield or topic, we only use these as a means of searching, not description for the clusters (i.e. to see where these subjects are distributed).

Patent categories

  • Each patent cluster is associated with a main category and a set of top categories — broad technology areas such as biotechnology, telecommunications, or mining and quarrying — determined using the categories of the individual patent families in the cluster.
    • Patent categories are drawn from a CSET classification scheme with two hierarchical levels: a broad category and a more specific technology category. Each patent family is assigned to one or more of these categories.
    • To determine a cluster's categories, we count how many patent families in the cluster fall into each category and select the three most common as the cluster's top categories, along with the percentage of the cluster's families associated with each. The single most common category is designated the cluster's main category.

Key concepts

  • Each cluster has a list of 5 key concepts, identified by applying a phrase-extraction algorithm to the articles in the cluster.
  • To identify key concepts in clusters, we used the yake algorithm to identify the top twenty keyphrases for each cluster based on their titles and abstracts, aiming for keyphrases that were relatively distinct from each other so that we would be able to identify discrete concepts. We then removed any phrases that were common English words (identified using the wordfreq Python library) or were single words that were part of other phrases that had been surfaced for the same cluster, to ensure that phrases were more useful and variable. The top five scoring keywords of those that remain were then assigned to the cluster.

Author and Country affiliation

Patent inventor and assignee

  • Top inventors are the inventors credited on the most patent families in the cluster.
  • Top assignees are the assignees (owners) — typically the companies, universities, or other organizations that hold a patent — credited on the most patent families in the cluster.
    • We use ROR data to assign a type to each assignee, which lets us separately surface a cluster's top academic and top industry assignees.

Maintenance

How are the data updated?

Clustering updates recalculate the clustering structure of the entire dataset "from scratch" using the most recent data and the method described above. The last clustering update was in December 2025.

Existing cluster metadata updates run roughly monthly.

Credits

Virtually all CSET data team members have contributed to the Clusters Dataset in some form, whether by providing feedback, reviewing code, or helping generate ideas. In alphabetical order, some particular contributions follow:

  • Rebecca Gelles: Phrase extraction and patent metrics development
  • Jennifer Melot: Clustering pipeline implementation and automation support; metadata aggregation and development support
  • Katherine Quinn: Clustering methodology and experimentation, metadata aggregation, ongoing metric development
  • Ilya Rahkovsky: Clustering methodology and experimentation, metadata aggregation, and primary metrics development
  • Autumn Toney: Metrics development and review

Dewey Murdick and SciTech Strategies helped develop the clustering concept and methodology. Katherine Quinn implemented structural changes to clustering concept and methodology to include citation and text similarity. Zach Arnold wrote the original documentation, and this updated version was written by Katherine Quinn, Adrian Thinnyun, and Catherine Aiken.

Emerging technology topic classifications are based on work supported in part by the Alfred P. Sloan Foundation under Grant No. G-2023-22358. Cluster titles and summaries are based on work supported by Google.org.

Major change log

10/13/22Initial release (ETO/CSET internal)
11/22/232023 update following changes to MAC (ETO/CSET internal)
12/15/25Changes to Clustering Method to include both citation and text similarity
4/27/26Cluster titles and text summaries added
7/16/26New Cluster Dataset published

Keep in touch

Twitter
LinkedIn
Substack
Email
RSS
Terms of Use and Privacy Policy