Abstract

Group profiling methods aim to construct a descriptive profile for communities in social networks. This task is similar to the traditional cluster labeling task, commonly adopted in document clustering to identify tags which characterize each derived cluster. This similarity encourages the direct application of cluster labeling methods for group profiling problems. However, in group profiling, an important additional information can be leveraged, which is the presence of links among the clustered individuals. This work extends our previous work by incorporating relational information to better describe communities. The proposed approach, so-called Centrality-based Group Profiling approach, makes use of network centrality measures in the selection of nodes for the characterization, i.e., nodes that generalize the content of the observed communities. The use of relational information to select relevant nodes in a community significantly reduces the complexity of the profiling task, at the same time retaining enough representative content to produce a good characterization. Experiments were conducted in a co-authorship network to evaluate different profiling strategies. The results demonstrated the ability of the proposed approach to producing good profiles for the observed groups with both group profiling and standard cluster labeling methods, with a considerably lower computational cost.

BibTeX

 @Article{Gomes2018,
  author="Gomes, Jo{\~a}o E. A.
  and Prud{\^e}ncio, Ricardo B. C.
  and Nascimento, Andr{\'e} C. A.",
  title="Centrality-Based Group Profiling: A Comparative Study in Co-authorship Networks",
  journal="New Generation Computing",
  year="2018",
  month="Jan",
  day="01",
  volume="36",
  number="1",
  pages="59--89",
  issn="1882-7055",
  doi="10.1007/s00354-017-0028-9",
  url="https://doi.org/10.1007/s00354-017-0028-9"
  }