Trees and Crops for the Future - international publications 2013-2024

Ver. 2 - Feb. 2025

1. Introduction

1.1 Indicators

Two indicators are used to highlight citation impact and international collaboration, but many more are availiable when downloading the data. Here we focus on:

• Top 10%: The proportion of publications that belong to the top 10% most cited in the world (database), taking into account publication year, publication type, and subject area. The expected value is, by definition, 10%.

• International Collaboration: The proportion of publications resulting from international collaborations. A value of, say, 40% means that this percentage of publications has at least two listed affiliations from different countries.

1.2 Implementation Aspects

The source of the above indicators is SciVal (based on Scopus data). When "subject areas" are used in the normalization process of the Top 10%-indicator described above, it refers to All Science Journal Classification (ASJC) categories. This classification system operates at the journal level. It is used here to provide a relatively coarse-grained subject normalization of the indicator. Furthermore, the observed values of the indicators are usually supplemented with so-called stability intervals (SI). These intervals are created by repeatedly and randomly selecting 90% of the publications without replacement and calculating the indicator value for the given subset. A 90% stability interval is then constructed by setting the lower (upper) limit to the 5th (95th) percentile in the resulting distribution.

When 'organizational units' constitute a variable, it is important to note that such data in its original form is not standardized. Throughout these exercises, organization disambiguation algorithms are used for identification and standardization. These are provided by a third party (Elsevier), and no manual corrections have been attempted. As is generally the case for bibliometric exercises, the focus should be on the overall structure that emerges. Further information about these disambiguation systems can be found here and here.

A note on the citation window: as the considered publication period is 2013-2024, citation metrics (in contrast to collaboration metrics, for example) for publications published during the last year convey little information. The potential number of received citations is usually very low, and we mostly model noise with an indicator like Top 10%. In addition, 2024 is not a complete year in a bibliographic sense at the start of 2025 (some publications are not yet indexed or might change the final publication year as the publication progresses through publication stages). All in all, interpret information for the latest year with caution or ignore it altogether.

1.3 Subject Partitioning^c

Even though the main level of analysis operates on the aggregated level of the full publication set, a tool for investigating publication volume, citation impact, and collaboration by subject matter is also available. Briefly, we cluster the publications based on their pairwise similarities, where similarity is based on the cited references and word usage in the publications (so-called "hybrid similarity"; see more here). The goal is to highlight the publication population's different research topics or themes. A technical description of the approach is available here and here.

After the publications have been clustered with respect to subject matter, overall descriptions or labels of the clusters are automatically created by analyzing the publications on a cluster-by-cluster basis with a large language model. Also, ASJC codes that are representative of the given cluster (but less so for others) and examples of frequent authors are extracted to help interpret the cluster's subject composition.

Note that an approach like this, which is entirely data-driven and combines bibliometric-based operationalization of subject similarity with machine learning and large language models, presents one view of the publication set. It should be clear that subject partitioning derived like this is naturally and necessarily partly a result of the current data model (how similarity between publications is operationalized), the choice of clustering method, and its associated free parameters (that, e.g., guide the resolution of the partition). While keeping this in mind, it can be a powerful analytic tool to help answer ad-hoc questions regarding the publication set.

The 752 publications have been grouped into 13 clusters.

2. Time Series - publication, citation impact, and collaboration

3. Subject Partitioning

Figure 4. Switch to full-screen mode by clicking the square in the upper right corner. Click on a node (publication) to get bibliographic information as well as Field-weighted Citation Impact (FWCI) and a link to the bibliographic record in Scopus. The control panel can be accessed by clicking the arrow found on the left edge. Zoom in and out using the buttons in the middle right (or use the scroll wheel on your mouse).

Table 1 complements the visualization in Figure 4.

Table 1. Subject partition - 13 clusters. Label, subfield, summary and keywords are extracted with the help of a Large Langue Model without manual intervention.

4. Collaborative Units

Figure 5. Switch to full-screen mode by clicking the square in the upper right corner. Click on a node (university/institution) show the number of publications attributed to the unit.

Table 2 complements the visualization in Figure 4.

Table 2. Units with at least 15 publications. Grouped by collaborative pattern (clustering).

Unit	Cluster	PUB
Sveriges lantbruksuniversitet (Sweden)	1	627
Skogforsk (Sweden)	1	63
Stockholms universitet (Sweden)	1	23
Commonwealth Scientific and Industrial Research Organisation (Australia)	1	23
Natural Resources Institute Finland (Luke) (Finland)	1	23
Duke University (United States)	1	23
Københavns Universitet (Denmark)	1	22
Helsingin Yliopisto (Finland)	1	21
Göteborgs Universitet (Sweden)	1	20
Asian School of the Environment (Singapore)	1	18
Umeå Universitet (Sweden)	2	193
Umeå Plant Science Centre (Sweden)	2	47
Beijing Forestry University (China)	2	31
The University of British Columbia (Canada)	2	24
Uppsala Universitet (Sweden)	2	21
Universiteit Gent (Belgium)	2	21
Norges Miljø- og Biovitenskapelige Universitet (Norway)	2	15
The Royal Institute of Technology (KTH) (Sweden)	3	60
Lunds Universitet (Sweden)	3	36
INNVENTIA (Sweden)	3	20

---

a. The following publication types are considered in all analyses: Article, Conference Paper, Review, Chapter, Book, Short Survey, and Data Paper. The following types are not included in the data: Editorial, Erratum, Retracted, Book reviews, and Abstract Report/Conference meeting abstracts.

b. See the Scopus Content Coverage Guide for a detailed description of Scopus content.

c. In the utilized clustering method ("multiresolution modularity" maximized with the so-called Leiden algorithm), γ controls the resolution ("number of clusters") of the partitioning. The problem now is to identify one or several reasonable values for γ. Here, we have followed an approach that creates a range of cluster solutions by systematically varying the value of γ and identifying regions where the cluster solutions are "stable," i.e., where small or no differences between the cluster solutions exist (Lambiotte, 2010). Stability is taken as a signal that the cluster solution is "robust" and, therefore, likely more interesting to focus on than cluster solutions sensitive to small changes in γ. It is easy to see that a hierarchy of cluster solutions should reasonably exist from thematically coarse-grained to very specific. In practice, one or several cluster solutions are chosen based on a combination of the stability criterion and a more subjective criterion stipulating that the solution should be "useful" for the purpose for which it was developed (here, essentially, "not too low resolution (few clusters) but not so high resolution that the benefit of automatically summarizing the publication volume is lost").

Cristian Colliander & Cecilia Sandberg @ UmUB - 2025-02-26.