Exploring characteristics of highly cited authors according to citation location and content

0.134 0.151 0.151 0.142 Background 0.133 0.134 0.133 0.133 Related work 0.151 0.153 0.153 0.148 Method 0.132 0.140 0.137 0.131 Evaluation 0.130 0.127 0.128 0.125 Discussion 0.132 0.133 0.133 0.127 Conclusion 0.130 0.135 0.132 0.126 For most of locations, the highest similarity value (bolded) is achieved when the window size of the context is 1. There is an exception, where in the case of Evaluation sections, the highest value occurs when all sentences in citation context are considered as boundary of citation context, but the difference between the optimal value and a window size of 2 is small. As such, we have uniformly applied a window size of 1 (i.e., the sentence hosting the citation expression, as well as the sentences directly before and after, when applicable) for all document sections in the remainder of this work. Content analysis using citation context To explore the characteristics of highly-cited authors in each section at a macro level, we calculated similarity among citation contexts of the top 100 cited authors by document section. We take the simplifying assumption that all authors of a paper contribute equally towards the paper in our work. We then applied the aforementioned JaroWinkler distance algorithm to calculate similarity of selected authors’ citation contexts as illustrated on the left side of Figure 6. The calculated similarity between citation contexts belonging to different authors are used (after summing) as weights for edges between the said authors in building an author network. We built an author network for each of the eight sections to compare distinct characteristics of author network according to the cited location in a document. FIG. 6. Workflow to construct author networks The resultant networks’ nodes are authors and edges reflect the similarity of citation contexts. Edge strengths are weighted by the similarities between citation contexts of the authors. To visualize the author networks, we employed Gephi 0.8.2. (Bastian et al., 2009), setting the size of (author) nodes relative to their degree centrality, and coloring the nodes to show the authors’ induced community as computed using the modularity algorithm (Blondel et al., 2008), with resolution set to 0.9. In other words, if the two nodes have the same color in the visualization, they have high similarity in their citation contexts. To expedite the analysis of each community, we also annotated the sub-discipline of the community based on the authors being included in each community.