Identifying Overlapping and Hierarchical Thematic Structures in Networks of Scholarly Papers: A Comparison of Three Approaches

We implemented three recently proposed approaches to the identification of overlapping and hierarchical substructures in graphs and applied the corresponding algorithms to a network of 492 information-science papers coupled via their cited sources. The thematic substructures obtained and overlaps produced by the three hierarchical cluster algorithms were compared to a content-based categorisation, which we based on the interpretation of titles and keywords. We defined sets of papers dealing with three topics located on different levels of aggregation: h-index, webometrics, and bibliometrics. We identified these topics with branches in the dendrograms produced by the three cluster algorithms and compared the overlapping topics they detected with one another and with the three pre-defined paper sets. We discuss the advantages and drawbacks of applying the three approaches to paper networks in research fields.

💡 Research Summary

This paper investigates how to uncover overlapping and hierarchical thematic structures in scholarly paper networks by implementing and comparing three recently proposed community‑detection algorithms. The authors apply the methods to a bibliographically coupled network of 492 information‑science articles published in 2008. To evaluate the results, they manually construct three reference topic sets—h‑index, webometrics, and bibliometrics—based on titles, abstracts, and keywords.
The first approach is the Local Fitness Maximization (LFM) algorithm introduced by Lancichinetti, Fortunato, and Kertesz. It defines a local fitness function f(C,α)=k_in(C)/k(C)^α, where α is a resolution parameter that controls the granularity of detected communities. By varying α, the algorithm yields a hierarchy of “natural” communities that can overlap because each node is examined independently using only its immediate neighborhood.
The second approach is hierarchical link clustering (HLC) as proposed by Ahn, Bagrow, and Lehmann. Instead of clustering nodes, it clusters citation links based on the Jaccard similarity of the neighborhoods of the two incident nodes. Each hard link cluster induces overlapping node communities: a paper’s membership grade in a community equals the proportion of its outgoing citation links that belong to the corresponding link cluster. This method exploits the presumed thematic homogeneity of individual citations.
The third approach starts from any hard‑clustering algorithm that produces disjoint community cores and then “fuzzifies” the boundaries. Nodes at the periphery are assigned fractional membership values µ_i(C) proportional to the weighted sum of their internal and external connections, allowing a node to belong partially to multiple communities.
The authors first illustrate each method on the classic Zachary karate‑club network, showing how overlapping and hierarchical structures emerge. They then apply the three algorithms to the 492‑paper network, generate dendrograms, and locate branches that correspond to the three manually defined topics. Comparative analysis reveals distinct strengths and weaknesses: LFM captures fine‑grained topics when α is high but is sensitive to parameter choice; HLC clearly reveals overlap through citation link homogeneity but can suffer when citation links are sparse or serve multiple purposes; fuzzification preserves the crisp core structure while improving recall, yet the interpretation of fractional memberships requires subjective thresholds.
Overall, the study demonstrates that overlapping, hierarchical thematic structures can be extracted from scholarly networks using purely local information, and that the three approaches provide complementary perspectives. The authors suggest that future work should explore scalability to larger corpora and integration with text‑based topic models to further validate and enrich the detected structures.