Source-linked AI summary
Interactive Overlays: A New Method for Generating Global Journal Maps from Web-of-Science Data
Loet Leydesdorff, Ismael Rafols
TL;DR
The paper addresses how to construct a global, interactive journal map while managing uncertainty from clustering, similarity, and visualization choices. It extends an earlier overlay using aggregated citation relations among 9,162 journals and compares alternative mapping procedures. The authors conclude that VOSViewer provides a suitable platform for interactive journal overlays, while emphasizing that methodological choices remain consequential.
Problem
Earlier global journal mapping faced unconvincing clustering, concern over similarity criteria, and solutions affected by parameter choices or random seeds.
Method
The study builds an interactive journal overlay from aggregated citation relations and compares cited versus citing data, MDS-like versus spring-embedded positioning, VOSViewer versus Gephi, and clustering choices.
Results
VOSViewer can serve as a platform for interactive journal overlays, with no major problems identified for this purpose.
Takeaways & Limitations
The overlay provides a way to position journals relative to larger journal structures while leaving some choices to users and supplying default options.
Takeaways & Limitations
Interpretation remains constrained by method choices, including similarity criteria and possible fluctuations in dynamic maps that may require significance testing.
Abstract
from arXiv · showhide
Recent advances in methods and techniques enable us to develop an interactive overlay to the global map of science based on aggregated citation relations among the 9,162 journals contained in the Science Citation Index and Social Science Citation Index 2009 combined. The resulting mapping is provided by VOSViewer. We first discuss the pros and cons of the various options: cited versus citing, multidimensional scaling versus spring-embedded algorithms, VOSViewer versus Gephi, and the various clustering algorithms and similarity criteria. Our approach focuses on the positions of journals in the multidimensional space spanned by the aggregated journal-journal citations. A number of choices can be left to the user, but we provide default options reflecting our preferences. Some examples are also provided; for example, the potential of using this technique to assess the interdisciplinarity of organizations and/or document sets.
1. Introduction
The paper develops an interactive journal-level overlay of global science to address limits in earlier local maps and unstable clustering or visualization choices. It motivates journal maps as useful for examining specialties, interdisciplinarity, and emerging fields.
- Contribution: The study extends an earlier interactive overlay to generate a global map of science at the journal level.The journal-level map is intended as an equivalent extension of the earlier subject-category overlay.
- Motivation: Earlier science maps were technically limited to local disciplinary or specialty views rather than comprehensive global journal maps.Journal-level mapping became possible as citation data and computational tools advanced.
- Scientific organization: Journal maps can identify specialties as relatively robust niches despite variation in precise clustering algorithms.Neighbouring specialties may remain weakly connected across intellectual borderlines and develop heterarchically rather than in an ordered hierarchy.
- Applications: Interactive visual-analytics tools can support policy-oriented identification of emerging fields or technologies at early stages.The paper links these uses to journal-level developments and the organization of scientific communication.
- Limitations: Global maps may appear robust at large scale, but lower-level journal structures can depend on small parameter changes.Visualization also reduces a multidimensional space to two dimensions, while some community-finding solutions depend on randomly chosen seeds and cannot be reproduced unambiguously.
2. Methods and data
The study constructs a cosine-normalized journal citation matrix from the 2009 Journal Citation Reports and compares visualization options before producing a larger interactive journal overlay. It evaluates cited and citing dimensions, thresholded network components, VOSViewer, Gephi, and related clustering choices.
- Data preparation: The matrix is cosine-normalized in both the being-cited and citing directions before visualization.Thresholded matrices can be exported in formats readable by visualization programs.
- Thresholding: At cosine > 0.5, the largest component contains 5,954 nodes, or 64.99%, connected by 51,030 edges, and is used first to compare methodological options.The smaller matrix helps assess visualization software before returning to the larger journal set.
- Visualization: VOSViewer handled the large component without an error message on an 8 GB, 64-bit machine, while Gephi required a long processing time for the same large file.The comparison concerns the component of 8,817 journals at cosine > 0.2.
- Visualization: VOSViewer uses an MDS-like algorithm that minimizes stress at the systems level, whereas Gephi provides graph-analytical routines and a spring embedder.VOSViewer also integrates clustering based on similar principles and allows its cluster parameter γ to be changed interactively.
3. Construction of the basemap in VOSViewer or Gephi?
The basemap combines journal-citation similarities with alternative visualization and clustering choices. VOSViewer and Gephi expose different trade-offs between scale, community separation, cross-community relations, and label readability.
- VOSViewer: The 8,817-journal set produced a memory error, while increasing Java memory to 8 GB and 64 bits resolved the problem.The authors therefore suggest basemap generation will soon be accessible to an average user.
- VOSViewer: VOSViewer generated a basemap for 5,954 journals with being-cited similarity above cosine > 0.5.The map displays 70 differently colored clusters and suggests broad axes separating social from natural and life sciences.
- Limitations: The community-finding algorithm can fluctuate because of random seeds and year-to-year changes in citation intensity, complicating comparisons across years.The paper cautions that visible changes in dynamic maps may occur without significance testing.
- Gephi and GEXFExplorer: Large overlays create practical readability limits: labels are too small in the overall map, cluttered at larger scales, and difficult to distinguish when close together.Lowering the cosine threshold can reduce cross-set problems but aggravates label overlap and size-related readability issues.
- Gephi and GEXFExplorer: ForceAtlas separates communities sharply and preserves intergroup linkages, whereas Fruchterman-Reingold provides a more compact representation.The layouts require choosing between sharper subset distinctions and displaying more relations across components.
4. The Cited Map as a baseline
The cited map uses journal being-cited similarities to provide a baseline at lower connectedness thresholds. It reveals broad scientific domains and bridges while reducing the number of displayed communities to a journal-level scale.
- Map structure: The bio-medical and natural-science domains are connected by biological and environmental sciences at the top and social sciences at the bottom.Computer-science and statistics journals appear near the center, while mathematics is set apart to the right.
- Interpretation: The reduction to 21 clusters is convenient because the objective is to map journals rather than hundreds of specialties.The map retains a torus-like structure comparable to other journal and subject-category maps.
- Interactive use: Journal coordinates can be saved in a comma-separated database file to connect Web-of-Science downloads with the map.Because abbreviations are not fully standardized between JCR and WoS, matching uses full journal titles.
5. The Citing Map as a baseline
The citing map positions journals by their citing patterns, offering a nearly complete but less decomposable network whose clustering and coloring can be changed independently. It provides a baseline in which citing behavior reaches across specialty boundaries, while cited patterns more regularly reproduce intellectual boundaries.
- Citing versus cited: Users may choose citing or cited patterns according to their research question: citing represents the knowledge base of new claims, whereas cited represents accumulated journal impact.The authors associate citing with continuously updated community activity and cited patterns with the scientific archive.
- Clustering: VOSViewer identifies 19 citing-pattern clusters at γ = 1, but only the first ten are considered meaningful because several clusters contain very few journals.Five clusters contain one journal and one contains four journals; at higher γ values, the number of clusters increases to 77 and 121.
- Network structure: Citing patterns are less decomposable than cited patterns, with authors reaching across specialty boundaries.The cited network more regularly reproduces boundaries in the intellectual organization of the sciences.
- Clustering: The mapping and clustering are uncoupled, allowing users to change the clustering and coloring while retaining the basemap.This separation is intended to reduce erroneous interpretations arising from dimensional reduction or clustering aggregation.
5. The generation of the overlay files
The overlay files convert Web of Science data into inputs that can be visualized and customized in VOSViewer. They support alternative classifications, multiple views, and adjustable node-size relations, but VOSViewer’s obligatory normalization limits quantitative comparison across overlays.
- File structure: Overlay files retain map coordinates and JCR journal titles, using journal titles as matching keys.An unforeseen title mismatch requires adapting the corresponding table entry.
- File generation: The programs generate either cited.txt or citing.txt from downloaded Web of Science data, and these files can be visualized in VOSViewer.The files are produced by separate routines and used as VOSViewer inputs.
- Visualization: VOSViewer provides label, density, and scatter views, along with controls for label visibility and colors.The overlay colors correspond to those in the global map.
- Node sizing: Node sizes default to log10(n + 1), preventing single-publication nodes from disappearing, but VOSViewer’s obligatory relative-size normalization complicates quantitative comparison across overlays.Linear publication-volume scaling can be obtained by replacing the weight values, yet animated maps remain unsuitable for direct visual comparison.
- Customization: Users can replace the default cluster classification while preserving the underlying map, including with the 41 communities identified by Blondel et al.’s algorithm.The mapping and clustering are therefore operationally separable.
6. An Application: Mapping “Interdisciplinarity”
The overlay method is applied to contested cases of interdisciplinarity in organizations and emerging specialties. The examples show that journal-map proximity and spread reveal disciplinary concentration, cross-disciplinary publication, and heterogeneity within broad categories.
- Application scope: The study tests overlay methods on interdisciplinary university units and emerging specialties such as nanoscience and nanotechnology.These applications address areas of contestation about how interdisciplinarity functions.
- Innovation Studies versus Business & Management: The overlay analysis provided quantitative evidence that evaluation tends to rank mono-disciplinary output more highly than interdisciplinary sets.This result is presented alongside the organizational comparison based on journal portfolios.
- Innovation Studies versus Business & Management: SPRU has a diverse publication profile, whereas London Business School is heavily concentrated in business and management journals.The comparison covers publication portfolios from 2006 to 2010.
- Innovation Studies versus Business & Management: The visualization conveys whether journals are proximate or distant in cognitive space, which tables and bar charts cannot show directly.It enables rapid comparison with the full journal set and provides a more precise representation than subject categories because individual journal names are mapped directly.
- Nanoscience and nanotechnology: Four factors from 58 of 59 journals explain 35.9% of the variance and divide the journals into four groups for coloring.Only journals with factor loadings larger than 0.4 on a factor were used for the Figure 8 coloring.
- Nanoscience and nanotechnology: In the nanoscience and nanotechnology map, journals associated with nanomedicine, microfluidic devices, chemistry and condensed matter physics, and electronics occupy distinct colored regions.The large distances between these groups raise questions about treating them as one interdisciplinary category.
Conclusions and discussion
The exploration supports VOSViewer as a platform for interactive journal overlays, while distinguishing its multidimensional approach from Gephi’s spring-embedded perspective and identifying remaining scaling limitations.
- VOSViewer as an interactive platform: VOSViewer can serve as a platform for interactive journal overlays.Users can retain default settings or supply their own preferences for colors, classifications, and node sizing.
- Remaining limitations: A drawback concerns scaling node weight relative to average weight, limiting control over node sizes and potentially affecting visual comparisons in animations.The authors suggest facilities like those available in Pajek as a possible improvement.
- Methodological comparison: Gephi was less successful in the exploration, with hurdles identified for possible resolution in future versions.Its spring-embedder was also considered conceptually less attractive because of the different topologies involved.
- Methodological comparison: VOSViewer positions journals in multidimensional space, whereas spring-embedded layouts emphasize individual network relations and topology.The authors describe VOSViewer as following the MDS tradition and using the cosine as a spatial measure suited to multidimensional mapping.
- Interpretive perspective: Graph relations indicate where information is communicated, while multidimensional positions specify communication from a systems perspective.The authors distinguish the network graph’s topology from the spatial representation of the broader system.
- VOSViewer as an interactive platform: The study finds no major problems with using VOSViewer for interactive overlays.The software also allows results to be uploaded on the Internet.