In the scientific community, literature resulting from academic research forms complex networks that include multiple information streams and huge data sets. As the keywords of literary works directly represent the main content or topics of research, we use keywords provided by authors to detect the hot topics in their fields. We can therefore discern the main focus of research fields and grasp the direction of their development, and such understanding can contribute to scientific advancement.
The term “hot topics” refers to subjects that attract extensive attention from researchers in a short period of time and was addressed frequently in literature (Wan et al., 2015). As the development of knowledge is continuous and fluid, changing focus in hot topics reflects the constant change in knowledge and information gathering. This idea drives us to find patterns in these changes and to detect hot topics from previous research in order to predict the future research direction.
Detecting dynamics of hot topics in a research field is central to discerning the core content of that research field. To detect hot topics based on scientific articles, researchers have employed various approaches such as word frequency analysis, co-word analysis, co-citation analysis, and social network analysis.
Word frequency analysis is a basic method of hot topic detection and often combined with other methods. For example, Gong and Ye (2006) analyzed research hotspots using index terms that captured the essence of a topic because they believed that keywords with a high frequency represented academic hotspots in a given field. Xiao (2011) employed the
Social network analysis (SNA), a popular approach at present, is a quantitative analysis approach based on mathematics and graph theory. It is able to build the social network model from complex literature networks and detect research hotspots, and it is widely used in Sociology, Information Science, Economics, and Management (Otte & Rousseau, 2002). Ding (2011) applied the topology- and topic-based community detection approaches to the coauthorship networks of information retrieval areas. She suggested that in the future the community detection approach should be used to identify dynamic changes of topics rather than emphasizing the relationships between communities and topics. Li et al. (2015) used classical word frequency analysis and co-word analysis along with centrality analysis and cohesive subgroups analysis to reveal the hot topics of international economic disciplines from 1999 to 2013. Song (2011) applied social network analysis to explore hot research topics and enhance the objectivity of measurements by drawing a global graph of co-citation networks and visualizing the graph’s components, bridges, cut-points,
Researchers have used different ways to track topic changes. Tang and Hu (2013) developed an integrative approach to tracking and visualizing the changes of research streams. Using research cohesion score, whose value is determined by the summation of shared keywords, they measured similarities in the focus of research of a pair of articles and found that the triggered research streams diffused via extended co-authorships.
With the development of visual technology in recent years, it is possible to have a more vivid display of hot topics’ dynamics. Many kinds of visualization software have been developed to deal with complex data. Among these visualization software, CiteSpace (Chen, 2006), VosViewer (van Eck & Waltman, 2010), and Sci2 (Sci2 Team, 2009) received the broadest acceptance. Hou et al. (2006) used knowledge maps instead of word frequency to identify research hotspots and trends of research fronts of international science studies. The alluvial diagram is an emerging visualization to reveal the process of mergers and splits of clusters over time. It was designed by Rosvall and Bergstrom in 2010 The alluvial diagram tool is available on
Publications in the research field of scientometrics were chosen for this study. Data was retrieved from Thomson Reuters’s (presently Clarivate Analytics) Web of Science (WoS) on May 16, 2016. To collect the most relevant publications in the field of scientometrics during the past five years, we conducted the literature retrieval in the following steps. First, we searched for articles published in
There are many popular timeline designs to fit different information streams, such as the three-dimensional spiral timeline, chessboard timeline, interaction timeline, relationship timeline, Gantt timeline, and complex timeline. Choosing an appropriate timeline approach depends on user demand. We used the alluvial diagram to demonstrate the dynamics of hot topics and to identify the structural changes of research. Specifically, main clusters in a scientific network at a given time occupy a column in the diagram and are horizontally connected to significant preceding and succeeding clusters by stream fields. We generated a co-word network before importing and generating the alluvial diagram. We then clustered the network and gave a label to each cluster. This approach relies on word profiles derived from articles citing a cluster of co-word articles, based on the assumption that the word profiles characterize the nature of a co-word cluster.
In the alluvial diagram, PageRank is used to reflect the importance of each cluster and word. We not only consider the frequency of keywords, but each keyword’s weight using PageRank (Page et al., 1998). PageRank is a link-analysis algorithm which is designed and used by Google to measure the importance of a webpage in the first place. After years of development, this algorithm has been used everywhere in network analysis. Using PageRank, we are able to identify hot topics more accurately than the traditional approach that only considers keyword frequency. PageRank gives each keyword a value of weight. The words used by a paper with high influence will be set a higher weight than those used by an ordinary paper. This idea was actually initiated from citation analysis (Pinski & Narin, 1976), which posits that the number of citations a paper receives can reflect the influence and importance of its research. PageRank algorithm extends this approach by not counting inbound links of all pages equally, but normalizing the number of outbound links and importance of neighboring pages, as shown in Equation (1):
where
We first calculated the frequency of keywords and chose the top 100 keywords with frequencies beyond 48. There were 28 significant keywords that needed to be normalized. Concentrating on hyphenated words and singular and plural nouns, we then replaced these keywords with normalized keywords in the original data set. This provided unitive data for further analysis.
After choosing the top 5% of frequent keywords from the articles, we generated five co-word networks with CiteSpace for years 2011 through 2015. By conducting cluster analysis, we detected hot topics for each year. After that, we drew an alluvial diagram between years, which displayed the dynamics of every cluster. In the alluvial diagram as shown in Figure 1, each block represents a cluster. The height of a block represents the cluster’s PageRank value. Blocks were ranked in descending order by PageRank value from the bottom to the top. To do this in a simple way, only the dynamics of the top 13 clusters with the highest PageRank value was shown, such as the cluster of
With the alluvial diagram in Figure 1, we can intuitively observe the mergers and splits of the top clusters from 2011 to 2015. The height of a stream field represents the flowing nodes’ total PageRank value. From this diagram, scientometrics has undergone constant development and changes in recent years. The
To analyze the importance of every cluster, we listed the name of every cluster and its PageRank value from 2011 to 2015 in Table 1.
PageRank values of the clusters from 2011 to 2015.2011 2012 2013 2014 2015 Clusters PR Clusters PR Clusters PR Clusters PR Clusters PR 1 #INDUSTRY 11.0% #PATENT CITATIONS 8.7% #TERMS 7.4% #INNOVATION 7.2% #FREQUENCY 7.3% 2 #GOOGLE SCHOLAR 10.0% #TRIPLE HELIX 7.3% #CITATION NETWORKS 7.3% #PUBLISHERS 5.8% #SEX DIFFERENCES 7.3% 3 #ITALY 8.2% #SCOPUS 6.9% #SCALES 6.8% #PERCENTILES 5.6% #SOCIAL NETWORKS 5.8% 4 #INDICATORS 8.2% #ACADEMIC WEB 6.8% #COOPERATION 6.4% #WEBOMETRICS 5.4% #CHINA 5.4% 5 #INTERNATIONAL COLLABORATION 7.5% #SCIENCE 6.2% #SCIENCE 5.5% #INFORMATION SCIENCE 5.4% #INFORMATION SCIENCE 5.2% 6 # 6.8% #CLASSIFICATION 5.9% #RESEARCH PRODUCTIVITY 5.4% #TWITTER 5.0% #TWITTER 5.1% 7 #RESEARCH PERFORMANCE 6.5% #SCIENTIFIC COLLABORATION 5.1% #INNOVATION 5.3% #ITALY 4.7% #CO-WORD ANALYSIS 5.0% 8 #RESEARCHERS 6.1% #MANAGEMENT 5.0% #RESEARCH PERFORMANCE 5.0% #RANKING 4.7% # 5.0% 9 #CO-WORD ANALYSIS 4.5% #PROXIMITY 4.7% #MAPS 4.7% #SCIENTIFIC PRODUCTIVITY 4.5% #SCOPUS 4.7% 10 #SOCIAL SCIENCES 4.4% #LIBRARY 4.7% #SCOPUS 4.6% #INTERDISCIPLINARY 4.5% #SELF-ORGANIZATION 4.6% 11 #CO-AUTHORSHIP NETWORK 4.2% #COOPERATION 4.7% #KNOWLEDGE 4.6% #SOCIOLOGY 4.3% #INN OVATION 4.6% 12 #RESEARCH COLLABORATION 3.9% #INDEX 4.4% #CO-AUTHORSHIP NETWORK 4.6% #PATTERNS 4.1% #CO-AUTHORSHIP 4.6% 13 #BIOTECHNOLOGY 3.7% #NANOTECHNOLOGY 4.2% # 4.3% #CO-AUTHORSHIP 4.0% #SCIENCE 4.5% 14 #WORD ANALYSIS 3.4% #DIFFUSION 4.1% #KNOWLEDGE MANAGEMENT 4.2% #NANOSCIENCE 3.9% #ITALY 4.2% 15 #INFORMATION SCIENCE 3.2% #SOCIAL NETWORK ANALYSIS 4.0% #CLASSIFIC ATION 4.0% #TRIPLE HELIX 3.6% #DEPARTMENTS 3.9%
As seen in Figure 1 and Table 1, we found that the cluster
The dynamics of the
The
In order to see the details of dynamic changes of the
.If the keyword is used frequently in 2011 as well as 2012, it will show up in the diagram. Every node represents a dynamic change, and there are no keywords with a high PageRank value during all five years. Although most keywords do show up in the following year, they do not have equal importance. In Figure 3, the keyword
After analyzing the dynamic changes in scientometrics from 2011 to 2015, we revealed the pattern of how a cluster was divided and merged over time. The topic could still be hot the following year, but other topics may have become more popular. As the alluvial diagram revealed, we were able to track a field of interest and which cluster it belonged to the following year. The changes in a research field reflect shifts in researchers’ interests and changing research objectives. Because scientometrics is a long-standing disciplinary field in Library and Information Science, a variety of methods can be applied to any discipline to solve research challenges and problems. While classic analytical methods continue to be important to scientometrics, methods coming from Information Science, which are more automated and efficient, are becoming more popular, as they can satisfy the increasing demands of big data processing.
Detecting the dynamics of hot topics means identifying their mergers and splits as research progresses. For example, the
Many researchers have applied various methods to find hot topics, including visualization technology that can highlight the hot fields. But as hot topics are only popular during a certain period of time, the time factor should be taken into consideration when gauging the topics value or influence. In networks, time is not easy to display clearly. The alluvial diagram can show the clusters clearly, while also adding the time slice to every cluster. By transforming the static network to a dynamic alluvial diagram, we can easily figure out the clusters’ contexts. Detecting the dynamics of hot topics in a field with the alluvial diagram can thus help researchers quickly ascertain the state of development of a field at the macro-level. Also, we can extract the keywords in each cluster. This allows us to better understand the internal knowledge structure of this field at the meso-level. Finally, we are able to track the changing path of the hot topics, which are reflected by the keywords, at the micro-level, and this can provide direction for researchers and suggestions for decision makers.
There are limitations to this study. The study period was only five years. When examining longer time periods, we encounter the problem of dealing with bigger data sets. Analyzing data year by year would be tedious, but if we combine e.g. two or more years into one time slice, important details would be missed. In our future studies, we aim to find a solution to this problem.