Academia.eduAcademia.edu

Figure 15 – uploaded by Chaomei Chen

See full PDFdownloadDownload figure
Browsing of individual articles confirmed that each cluster was comprised of articles where researchers used the technique central to the cluster. Identical queries were made to the Cartographic-SOM, co-term, and LSA maps, with the results shown in Figure 16 In these three maps, the query results are spread throughout the maps and are not associated with any discrete clusters. Thus, articles are clustered much better by technique when using references than when using term or title-based similarity measures. This is likely due to lack of consistent use of terms corresponding to techniques in either keyword lists or titles. We expect that maps based on similarities between words in abstracts would do a much better job of clustering by technique than do any of the measures shown here. Figure 16. Comparison of distinct fields on the citation map with their counterparts on the term or title-based maps. Legend — BLUE: author co-citation (ACA), GREEN -— co-word analysis and Leximappe, MAGENTA: co-citation analysis.

Figure 16 Browsing of individual articles confirmed that each cluster was comprised of articles where researchers used the technique central to the cluster. Identical queries were made to the Cartographic-SOM, co-term, and LSA maps, with the results shown in Figure 16 In these three maps, the query results are spread throughout the maps and are not associated with any discrete clusters. Thus, articles are clustered much better by technique when using references than when using term or title-based similarity measures. This is likely due to lack of consistent use of terms corresponding to techniques in either keyword lists or titles. We expect that maps based on similarities between words in abstracts would do a much better job of clustering by technique than do any of the measures shown here. Figure 16. Comparison of distinct fields on the citation map with their counterparts on the term or title-based maps. Legend — BLUE: author co-citation (ACA), GREEN -— co-word analysis and Leximappe, MAGENTA: co-citation analysis.

Related Figures (21)

Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the Unitec States Department of Energy under Contract DE-AC04-94AL85000. This is the table of contents of Katy Borner, Chaomei Chen, & Kevin Boyack: Visualizing Knowledge Domains. Annual Review of Information Science & Technology, Volume 37, 2003. (in press) Visualizing Knowledge Domains >
Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the Unitec States Department of Energy under Contract DE-AC04-94AL85000. This is the table of contents of Katy Borner, Chaomei Chen, & Kevin Boyack: Visualizing Knowledge Domains. Annual Review of Information Science & Technology, Volume 37, 2003. (in press) Visualizing Knowledge Domains >
Figure 1. Process flow for mapping knowledge domains. or the assignment of coordinates to each unit, and (6) use of the resulting visualization for analysis and interpretation. Steps four and five of this process are often distilled into one operation, which can be described as data layout.
Figure 1. Process flow for mapping knowledge domains. or the assignment of coordinates to each unit, and (6) use of the resulting visualization for analysis and interpretation. Steps four and five of this process are often distilled into one operation, which can be described as data layout.
Table 1: Comparison of Techniques One exception to Table lis noted here. Although VxOrd is classified as a FDP algorithm, it does not act like FDPs as characterized in the table. Rather, it can scale to very large data sets (millions of similarity pairs, only limited by memory constraints), has very fast run times, and provides a static layout. Some ordination techniques have a very high computational complexity for large data sets. This complexity can be reduced at the cost of global optimality by breaking the data set into smaller chunks, ordinating each cluster, and compiling all clusters into a single map.
Table 1: Comparison of Techniques One exception to Table lis noted here. Although VxOrd is classified as a FDP algorithm, it does not act like FDPs as characterized in the table. Rather, it can scale to very large data sets (millions of similarity pairs, only limited by memory constraints), has very fast run times, and provides a static layout. Some ordination techniques have a very high computational complexity for large data sets. This complexity can be reduced at the cost of global optimality by breaking the data set into smaller chunks, ordinating each cluster, and compiling all clusters into a single map.
Table 2. Search terms used to generate the ARIST bibliographic data set.
Table 2. Search terms used to generate the ARIST bibliographic data set.
several things; for instance, the journal may not require index terms from the authors, the journal may not supply terms to the database vendor, or the database vendor may choose not to index an article.) This makes the use of terms a less-than-optimum basis for mapping of these data. By contrast, abstracts are available for 84% of the post-1991 ARIST data set, making it a richer source of information. The percentages listed here apply only to this data set; we do not know the Book, journal and/or conference coverage can also be an issue. For instance, JASIS(T), Scientometrics, Journal of Information Science, Information Processing and Management, and the Journal of Documentation are key sources for visualization of science or knowledge domains (see Yet the SCI only started coverage of these journals in the mid-1990s. (JASIS, Scientometrics, and JDoc were covered through the mid-1980s.) Queries to the SCI alone would not have provided sufficient coverage of the intended fields over this time period due to this lack of key journal coverage. Thus, we queried the SSCI as well, which covered those journals over the entire time period under investigation. Users of ISI’s Web of Science can conduct a search across all databases such as the SCI, SSCI, and the Arts and Humanities Citation Index (AHCI).
several things; for instance, the journal may not require index terms from the authors, the journal may not supply terms to the database vendor, or the database vendor may choose not to index an article.) This makes the use of terms a less-than-optimum basis for mapping of these data. By contrast, abstracts are available for 84% of the post-1991 ARIST data set, making it a richer source of information. The percentages listed here apply only to this data set; we do not know the Book, journal and/or conference coverage can also be an issue. For instance, JASIS(T), Scientometrics, Journal of Information Science, Information Processing and Management, and the Journal of Documentation are key sources for visualization of science or knowledge domains (see Yet the SCI only started coverage of these journals in the mid-1990s. (JASIS, Scientometrics, and JDoc were covered through the mid-1980s.) Queries to the SCI alone would not have provided sufficient coverage of the intended fields over this time period due to this lack of key journal coverage. Thus, we queried the SSCI as well, which covered those journals over the entire time period under investigation. Users of ISI’s Web of Science can conduct a search across all databases such as the SCI, SSCI, and the Arts and Humanities Citation Index (AHCI).
lable 3. Number of articles by journal in the ARIST set (10 or more articles per journal). For completeness, we include a distribution of the number of articles per field per year, along with average citation counts, for the ARIST data set (see Figure 3). This shows the dramatic increase in publishing in citation analysis and bibliometrics starting in the late 1980s, and the birth of the visualization field around the same time. It also shows that citation analysis articles were more highly cited that bibliometrics articles in the 1970s and 1980s, and that citation counts for all four fields have generally dropped throughout the 1990s. The most recent articles have, of course, been cited infrequently due to their young age.
lable 3. Number of articles by journal in the ARIST set (10 or more articles per journal). For completeness, we include a distribution of the number of articles per field per year, along with average citation counts, for the ARIST data set (see Figure 3). This shows the dramatic increase in publishing in citation analysis and bibliometrics starting in the late 1980s, and the birth of the visualization field around the same time. It also shows that citation analysis articles were more highly cited that bibliometrics articles in the 1970s and 1980s, and that citation counts for all four fields have generally dropped throughout the 1990s. The most recent articles have, of course, been cited infrequently due to their young age.
Figure 3. Numbers of articles by field per year in the ARIST data set with average citation counts. Articles contribute to counts in more than one field if retrieved by queries from multiple fields.
Figure 3. Numbers of articles by field per year in the ARIST data set with average citation counts. Articles contribute to counts in more than one field if retrieved by queries from multiple fields.
The resultant author co-citation map contains 380 authors who have nine or more citations over the entire period between 1977 and 2001. Pathfinder network scaling limited the number of “salient” connections among these authors to 384. As usual, each author's node is colored by the factor loadings in the largest three specialties. An author co-citation map of a focused, coherent subject domain should demonstrate a considerable degree of conformance between the Pathfinder network structure and the factor analysis color patterns. However, this is not the case here, suggesting that science mapping constitutes a number of largely independent disciplines - as (Leydesdorff & Wouters, 2000) describes the status of scientometrics - this is pre- naradiomatic. Figure 4. An overview of the author co-citation map (1977-2001), consisting of 380 authors with 9 or more citations. The map is dominated by the largest specialty of citation indexing. No strong concentration of other specialties are found, which implies the diversity of the domain.
The resultant author co-citation map contains 380 authors who have nine or more citations over the entire period between 1977 and 2001. Pathfinder network scaling limited the number of “salient” connections among these authors to 384. As usual, each author's node is colored by the factor loadings in the largest three specialties. An author co-citation map of a focused, coherent subject domain should demonstrate a considerable degree of conformance between the Pathfinder network structure and the factor analysis color patterns. However, this is not the case here, suggesting that science mapping constitutes a number of largely independent disciplines - as (Leydesdorff & Wouters, 2000) describes the status of scientometrics - this is pre- naradiomatic. Figure 4. An overview of the author co-citation map (1977-2001), consisting of 380 authors with 9 or more citations. The map is dominated by the largest specialty of citation indexing. No strong concentration of other specialties are found, which implies the diversity of the domain.
It is interesting to note that none of the four largest factors are centered in either visualization or semantics, which were two of the four groups comprising the ARIST data set. These two fields are relatively new and are not highly cited (see Figure 3). Thus, they are unlikely to be strong factors in an ACA-type analysis.
It is interesting to note that none of the four largest factors are centered in either visualization or semantics, which were two of the four groups comprising the ARIST data set. These two fields are relatively new and are not highly cited (see Figure 3). Thus, they are unlikely to be strong factors in an ACA-type analysis.
Figure 6. An overview of the document co-citation map of 394 articles with 10 or more citations. In this network, a number of tight clusters of documents are connected to an artery-like chain. Documents on the artery chain tend to be seminal works of connected clusters. For example, Diana Crane’s Invisible College (Crane, 1972) connects the scholar communication cluster to the artery chain.
Figure 6. An overview of the document co-citation map of 394 articles with 10 or more citations. In this network, a number of tight clusters of documents are connected to an artery-like chain. Documents on the artery chain tend to be seminal works of connected clusters. For example, Diana Crane’s Invisible College (Crane, 1972) connects the scholar communication cluster to the artery chain.
Bin Zhu and Hsinchun Chen of the University of Arizona visualized the ARIST data set using ET-Maps (see section 4.3.2). They trained 10*10 nodes using the ID/keyword data of the ARIST data set. At the end of the map creation process, each node is associated with a list of documents that are semantically similar to each other. In addition, a phrase was assigned to each node as its label and adjacent nodes that have the same label are grouped into one region or category. Thus, spatial proximity on the map indicates semantic proximity, meaning if two categories are close to each other on SOM map, they are also semantically similar to each other.
Bin Zhu and Hsinchun Chen of the University of Arizona visualized the ARIST data set using ET-Maps (see section 4.3.2). They trained 10*10 nodes using the ID/keyword data of the ARIST data set. At the end of the map creation process, each node is associated with a list of documents that are semantically similar to each other. In addition, a phrase was assigned to each node as its label and adjacent nodes that have the same label are grouped into one region or category. Thus, spatial proximity on the map indicates semantic proximity, meaning if two categories are close to each other on SOM map, they are also semantically similar to each other.
listed on the left hand side together with a number in brackets telling how many individual documents it contains. In a typical browsing session a user would get an overview first, zoom into areas of interest, and access a particular document. Alternatively, a user can select a category of interest and the interface will display documents within that category. Figure 9. ET-Map of the ARIST data set using keywords. Left is a list of labels of all categories. A user can select a category of interest and the interface will display documents within that category.
listed on the left hand side together with a number in brackets telling how many individual documents it contains. In a typical browsing session a user would get an overview first, zoom into areas of interest, and access a particular document. Alternatively, a user can select a category of interest and the interface will display documents within that category. Figure 9. ET-Map of the ARIST data set using keywords. Left is a list of labels of all categories. A user can select a category of interest and the interface will display documents within that category.
Figure 10. Cartographic SOM map of ARIST data set. For each cluster, three label terms are computed, to better convey cluster content. Since the clusters were computed from the SOM itself, these labels indicate the potential topic or theme that one would expect to find with articles assigned to a particular cluster. In order to show the relative prominence of topics in the data set, the clusters were ranked according to the number of articles of this data set that were actually assigned to them. Clusters containing more articles appear larger and more prominent.
Figure 10. Cartographic SOM map of ARIST data set. For each cluster, three label terms are computed, to better convey cluster content. Since the clusters were computed from the SOM itself, these labels indicate the potential topic or theme that one would expect to find with articles assigned to a particular cluster. In order to show the relative prominence of topics in the data set, the clusters were ranked according to the number of articles of this data set that were actually assigned to them. Clusters containing more articles appear larger and more prominent.
The fourth segment (1995-2000) shows semantic analysis as a new field (labeled C), which, like the visualization clusters, is not well linked to the main body of work. The citation analysis and bibliometrics fields each seem to be well defined, with some mixing of the two in certain clusters. While bibliometrics seemed to be growing faster than citation analysis in the middle two segments (late 80s and early 90s), comparison of the numbers of papers in the late 90s shows citation analysis to be gaining strength. There is also an interesting cluster (labeled D) in which citation analysis, bibliometrics, and visualization are mixed. Additional analysis on this and surrounding clusters will be given in section 6.2 (see Figure 14).
The fourth segment (1995-2000) shows semantic analysis as a new field (labeled C), which, like the visualization clusters, is not well linked to the main body of work. The citation analysis and bibliometrics fields each seem to be well defined, with some mixing of the two in certain clusters. While bibliometrics seemed to be growing faster than citation analysis in the middle two segments (late 80s and early 90s), comparison of the numbers of papers in the late 90s shows citation analysis to be gaining strength. There is also an interesting cluster (labeled D) in which citation analysis, bibliometrics, and visualization are mixed. Additional analysis on this and surrounding clusters will be given in section 6.2 (see Figure 14).
Figure 12. VxInsight co-term (left) and LSA (right) maps of ARIST data The map based on LSA over article titles shows some different features. The titles contained 4,802 unique words, while abstracts contained 14,494 unique words. Here, the semantics papers are spread throughout the landscape. The visualization papers are also more mixed, and in several cases, they seem to bridge groups of citation analysis and bibliometrics articles. Bibliometrics and citation analysis articles, while appearing in the same clusters, are segregated within the clusters.
Figure 12. VxInsight co-term (left) and LSA (right) maps of ARIST data The map based on LSA over article titles shows some different features. The titles contained 4,802 unique words, while abstracts contained 14,494 unique words. Here, the semantics papers are spread throughout the landscape. The visualization papers are also more mixed, and in several cases, they seem to bridge groups of citation analysis and bibliometrics articles. Bibliometrics and citation analysis articles, while appearing in the same clusters, are segregated within the clusters.
Figure 13. VxInsight co-classification map of ARIST data
Figure 13. VxInsight co-classification map of ARIST data
Table 4. Comparison of document maps. \ second issue concerns the number of terms associated with each article. For the ARIST data, 34% of the articles with terms have 1-3 terms, 42% of the articles have 4-7 terms, 17% of the irticles have 8-10 terms, and the remaining 7% have 11 or more terms. Chung and Lee (Chung & Lee, 2001) showed that cosine measures tend to emphasize high-frequency terms. Andre skupin’s experience (see section 6.1.3) with SOM is that articles with only | or 2 terms tend to ither eqnsregate in less dense areas of the SOM or in the middle of clusters, rather than at their dges 22 Boyack’s experience using a cosine co-term similarity with FDP is that articles with mly 1 or 2 terms tend to be evenly distributed if the associated terms are specific, multi-word erms, but lie in Ve middle of large clusters if the associated terms are single word, general terms e.g, SCIENCE). Different results may arise from different layout routines; nevertheless, a lefinitive study on the effects of term types, the generality of terms, and term distributions is 1eeded.
Table 4. Comparison of document maps. \ second issue concerns the number of terms associated with each article. For the ARIST data, 34% of the articles with terms have 1-3 terms, 42% of the articles have 4-7 terms, 17% of the irticles have 8-10 terms, and the remaining 7% have 11 or more terms. Chung and Lee (Chung & Lee, 2001) showed that cosine measures tend to emphasize high-frequency terms. Andre skupin’s experience (see section 6.1.3) with SOM is that articles with only | or 2 terms tend to ither eqnsregate in less dense areas of the SOM or in the middle of clusters, rather than at their dges 22 Boyack’s experience using a cosine co-term similarity with FDP is that articles with mly 1 or 2 terms tend to be evenly distributed if the associated terms are specific, multi-word erms, but lie in Ve middle of large clusters if the associated terms are single word, general terms e.g, SCIENCE). Different results may arise from different layout routines; nevertheless, a lefinitive study on the effects of term types, the generality of terms, and term distributions is 1eeded.
Figure 14. Comparison of layouts of four different document maps based on terms or words. A: Cartographic-SOM (compare [Figure 10), B: ET-Map (compare Figure 9), C: Co-term (compare Figure 12, left), D: LSA (compare Figure 12, right). Dot color legend — YELLOW: citation analysis, GREEN: bibliometrics, BLUE: semantics, MAGENTA: visualization. level, the different layout techniques seem to give similar groupings from the term or word expressions. We have not done a neighborhood analysis to verify this statistically, but leave that to another time and study.
Figure 14. Comparison of layouts of four different document maps based on terms or words. A: Cartographic-SOM (compare [Figure 10), B: ET-Map (compare Figure 9), C: Co-term (compare Figure 12, left), D: LSA (compare Figure 12, right). Dot color legend — YELLOW: citation analysis, GREEN: bibliometrics, BLUE: semantics, MAGENTA: visualization. level, the different layout techniques seem to give similar groupings from the term or word expressions. We have not done a neighborhood analysis to verify this statistically, but leave that to another time and study.
Figure 15. Strong co-term linkages based on cosine similarity for the three term-based document maps of Figure 14] perhaps fewer clusters and less efficient space-filling can be justified. The LSA map is not compared here since title words and terms give different levels of information (Qin, 2000).
Figure 15. Strong co-term linkages based on cosine similarity for the three term-based document maps of Figure 14] perhaps fewer clusters and less efficient space-filling can be justified. The LSA map is not compared here since title words and terms give different levels of information (Qin, 2000).
Related topics:
Library and Information StudiesComputer science & Information TechnologyComputer Science Information Technology
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved