Analysis of Citation Networks - SergiuTripon/msc-thesis-na-epsrc GitHub Wiki
Home ▸ Literature Survey ▸ Analysis of Citation Networks
Year: N/A
Authors: Anita Valmarska, Janez Demšar
File: valmarska.pdf
Contents
- 1. Introduction
- 2. Data collection and network construction
- 3. Community detection and naming the communities
- 4. Results
- 5. Conclusion
1. Introduction
In the scientific world citations are used also to critically analyze or correct earlier work. Intuitively, it can be expected that papers would more often cite other papers from the same research subfield. To confirm this hypothesis, we wanted to find out whether we could detect the research subfields within a single research field, i.e., psychology, using only a citation network of papers published in the given research field.
2. Data collection and network construction
Consequently, we decided to crawl the pages connected with psychology in Wikipedia. From each of the visited pages we collected the references identified by their DOIs in the reference section. This resulted in a collection of 63,826 unique papers.
Next, we queried the Microsoft Academic Research data (MAS) and collected information about the scientific papers citing the initial set of collected papers. This allowed us to construct a citation network whose core contained papers published in the field of psychology. The resulting network consists of 948,791 vertices and 1,539,563 edges.
Due to the nature of our data collection process, we had to perform an initial data pre-processing in order to extract the papers that had a significant impact on the field of psychology. This resulted with a new network of 3,918 vertices connected by 5,732 edges.
3. Community detection and naming the communities
The process of identification of research subfields in the citation network was translated into the problem of community detection. For the purpose of our research, we applied the Louvain method. It is a simple, efficient, and easy to implement method for identifying communities in large networks.
Part of the evaluation of the detected communities was to name them and examine their connections. Due to the vast quantity of available data and unfamiliarity with the field of psychology, we named the communities based on the cosine similarity between our initially collected psychological papers and relevant texts for each of the APA (American Psychological Association) divisions of psychology.
4. Results
The community detection algorithm implemented in Pajek detected 52 communities. The smallest cluster included 7 papers, while the largest cluster was constructed of 230 psychological publications.
5. Conclusion
Results obtained by the network analysis and community detection are encouraging. The visual representation of the communities reveals sensible relationships between psychology subfields. However, the nature of data collection and the influence of our subjective judgment on community naming offer opportunities for further improvement. This involves improved data collection, developing new and improved methods for community detection, and employing better measures for text similarity. In further work, we would also like to explore the methodology proposed by Grčar et al.