Usage Count: A New Indicator to Detect Research Fronts

1 Introduction

Research fronts detection has become the focus of global scientific and technological competition. Through detecting and tracking research fronts timely and accurately, Japan (Kuwahara, 2007; Nagano, 2005) and the US (Porter, Guo, & Chiavatta, 2011) have made significant advancements in science and technology (S&T) policy-making and technological evaluation.

Research front was originally a term used by Price (1965). He concluded that a research front is a small part of earlier literature knitted together by the new year’s crop of paper, and used the phrases “epidermal layer” and “growing tip” to describe research fronts. Since then, scholars have made efforts to identify research fronts from various points of view. Small and Griffith (1974) considered co-citation clusters as research fronts; Vlachý (1984) summarized prior research on scientometric studies on research fronts detection and pointed out that “science grows from a very thin skin of its research front” and “a core body of seminal literature” constitutes “a sort of epidermal layer, an active research front” (p. 95). Garfield (1994) pointed out that research fronts are co-citation clusters plus citing articles; Morris et al. (2003) applied bibliographic coupling methods to identify the research fronts; Shibata et al. (2008) proposed that research fronts are direct citation clusters. Presently, Chen (2006), Braam, Moed, and van Raan (1991), and Persson’s (1994) views are the mainstream in research fronts detection. They concurred that research fronts are clusters of citing papers sharing a common intellectual base.

The research fronts detection method of this paper is in accordance with Chen, Braam et al., and Persson’s points of view. We label groups of citing articles that cite clusters of co-cited references as research fronts, and they are labeled from the title of an article which cites the most references in the cluster. When beginning co-citation analysis, the usual method is first to set a threshold and find the representative highly cited papers by “times cited,” and then make a co-cited matrix before clustering the networks and identifying the research fronts. However, due to the large time lag, there is a problem in using times cited as an indicator in research fronts identification. It might take up to two years for a paper to become highly cited (Shibata et al., 2008), and the situation varies among disciplines. Besides, times cited is affected by authors’ different citing motivations, articles accessibility (Bollen et al., 2005), etc. Therefore, as a traditional indicator, times cited cannot reflect the current interests of the research community. Faced with the fast paced development of S&T, new methods, tools, and indicators need to be developed to capture the research fronts more precisely in order to support S&T policy-making.

Open access journal publishers such as PLoS^[①] took the lead to provide online usage data for published articles. When more and more publishers chose to disclose the usage data of academic articles, researchers proposed to use articles’ usage data as potential complements, perhaps even alternatives, for research evaluation (Das & Mishra, 2014; Yan & Gerstein, 2011). They emphasized the superiority of usage data to citation data, such as ease to access and more convenient data collection. However, there are divergent opinions with regard to whether usage data such as times of views and downloads can be used as metrics of research evaluation. Some researchers reported significant correlation between specific usage types (Line & Sandison, 1975), especially downloads and citations, but others (Schloegl & Gorraiz, 2010; 2011) found only moderate or a rather low correlation between downloads and citations.

On September 26, 2015, Thomson Reuters started to provide article-level usage data, called “usage count” on the Web of Science (WoS) platform (Thomson Reuters, 2015). The new indicator, consisting of “U1” and “U2,” reflects the user’s level of interest by marking an article when read or downloaded by researchers. U1 is the count of the number of times the full text of a record has been accessed or saved within the last 180 days. U2 is the count of the number of times the full text of a record has been accessed or saved since February 1, 2013. The usage count is recorded every second day after users full-text request of an article, exports to bibliographic management tools or to formats for later import into bibliographic management tools, thus it does not need to wait for the tedious submitting and publishing process of times cited.

There is limited research on the effects of usage data on research evaluation. Presently, we have retrieved two relative articles (Martín-martín, 2016; Wang, Fang, & Sun, 2016) involving the relationship between WoS usage count and times cited. They discussed the usage patterns of articles and the correlation between usage count and times cited. In their research, Wang, Fang, and Sun (2016) discovered that citations play an important role in determining the usage count for old papers, and highly cited old papers are more likely to be used for even a long time after publication. Following their study, we are curious whether usage count^[②] can become a new and more effective indicator in detection of the latest research fronts, which are valuable for decision-makers to track the latest research trend and take the lead in scientific competition. To compare research fronts detected based on different indicators, we use “recentness” as a measure and we define recentness of a research front as the average publication year of citing papers of the co-citation cluster. Our research questions are:

1. What is the difference in recentness of research fronts generated by usage count and times cited?

2. Are there any common research fronts detected by usage count and times cited, and if there are, what about the recentness?

3. What is the difference in recentness of the top 10 highly cited papers selected by usage count and times cited?

2 Data and Methodology

2.2 Methodology

Co-citation analysis is the typical bibliometric method, initially proposed by Small and Griffith (1974). Articles were clustered together based on their co-occurrence in the references lists of papers. In other words, if articles A and B are both cited by article C, it is more likely that they belong to the same research field and share similar topics or methods.

To identify clusters which represent the intellectual bases, spectral clustering algorithms have been used in this paper. Spectral clustering is a clustering method that uses eigenvectors of an affinity matrix derived from the data (Dhillon, 2004), and results derived by spectral clustering often outperform the traditional algorithms such as k-means or single linkage (von Luxburg, 2007). Besides, spectral clustering is easy to implement. This study uses CiteSpace 4.0.R5 (Chen, 2006) to identify the co-citation clusters and furthermore to find the citing articles of each cluster as well as the research front.

There are 84,315 and 113,339 references in dataset “times cited” and dataset “usage count,” respectively. In order to eliminate records that have little relationship with our research and pick out the most frequently used references, we select the top 1% most-cited records within the datasets to be further analyzed. In order to reduce disparities caused by absolute frequencies, we construct the co-citation matrix in terms of cosine coefficients. Additionally, a minimum spanning tree network is used in the software for network pruning in order to hide relatively weak citation links between item i and item j in the matrix, and improve the pruning efficiency as well. We compare the average publication year of citing articles (the recentness) as well as the mean cited year of the co-citation clusters. In this paper, average publication year of citing articles is equal to the mean publication year of the citing publications, and does not take the number of citations into account. Figure 1 presents organization of this study.

Figure 1

Organization and procedure of the present study. RF is the abbreviation for research front and TC for times cited. UC refers to usage count.

3 Results

3.1 Basic Statistic Comparison of Times Cited and Usage Count

We investigate the distribution of citing articles and cited references within the two datasets. Figure 2 shows that the citing articles of usage count have experienced an exponential growth, because researchers prefer to use newly published literature, while Figure 3 indicates a right-skewed distribution for citing articles of times cited. This phenomenon is an indication that it will take several years for a paper to become highly cited. The distribution of cited references in Figures 4 and 5 presents a right-skewed distribution, but the data collected by usage count present a relatively high real-time property when compared with that collected by times cited. Many cited references were gathered from 2010 to 2013 by usage count in comparison with 2005–2008 by times cited.

Figure 2

Yearly frequency distribution of citing articles with threshold of usage count. Y represents the usage count and R² reflects the goodness-of-fit of this equation. A value closer to 1 indicates a satisfactory goodness-of-fit.

Figure 3

Yearly frequency distribution of citing articles with threshold of times cited.

Figure 4

Yearly frequency distribution of cited references with threshold of usage count. The time span of cited references is 1638–2015, and the study integrated the data of 1638–1990 to facilitate reading.

Figure 5

Yearly frequency distribution of cited references with threshold of times cited. The time span of cited references is 1632–2014, and the study integrated the data of 1632–1990 to facilitate reading.

There is approximately a four-year time lag in the mean publication year of citing articles, as well as a three-year time lag in the mean publication year of cited references collected by times cited. As is shown in Table 1, the recentness in times cited is 2009 and 2013.3 in usage count. As for the mean publication year of cited references, the figure is 2002.6 in times cited and 2005.7 in usage count.

Table 1

Overview of the dataset.

	Times cited	Usage count
No. of cited references	84,315	113,339
Recentness	2009.0	2013.3
Mean publication year of cited references	2002.6	2005.7

Note. Recentness means the average publication year of citing papers.

4 Discussion and Conclusion

The study collects 2,000 records by both times cited and usage count to measure whether usage count can be a new indicator in detection of research fronts. We find both indicators can be used in detection of research fronts, but using usage count can detect the latest research fronts than using times cited. In comparing the effects of the two indicators, first, we note that the majority of research fronts generated by usage count tend to be newer than times cited. Second, we investigate the recentness of a common research front detected by usage count and times cited. Results indicate using usage count can detect the latest research fronts than using times cited. Third, we compare the top 10 most highly cited papers detected by usage count and times cited. We find the top 10 papers selected by usage count represent more recent research fronts than selected by times cited. Moreover, we draw the conclusion that research fronts detected by usage count tend to be within the last two years, and present a higher immediacy and real time accuracy compared with times cited. Usage count can greatly shorten the time lag in research fronts detection, which could become a complementary indicator in the recentness detection of research fronts.

Usage count would be a new indicator in recentness detection of research fronts. If paper A is cited frequently within a period of time, the times cited will be added to WoS once the citing articles are published online. Usage count captures the researchers’ preference on various publications within the last 180 days. Generally, researchers prefer to use newly published papers, and therefore the usage data from publications within the last three years will reach a peak with relatively few citations (Wang, Fang, & Sun, 2016). Therefore, the meta data collected by usage count are most likely to be recent publications. In the research front detecting process, cited references are clustered as intellectual base and the citing articles form the “footprints” of research fronts accordingly. Citation activity can lag behind the publication of an article and some research domains are slow to be cited. In this sense, there is a relatively larger time lag in the research front detection based on times cited.

This paper represents preliminary work on the study of usage count in research fronts detection. However, there are some limitations in the study. For instance, the research fronts generated based on co-citations may refer to the hot research fronts, while we are trying to identify the cutting-edge research fronts. The usage count of older highly cited papers were not taken into consideration, because the new usage count indicator released by WoS only reflects usage logs after February 2013. In comparison to times cited, usage count is a dynamic and instant indicator. However, the correlation between usage count and times cited needs to be further discussed in the future.