Evolution of Co-Authors Communities Formed by Terms on DBLP

Alisa Babskova, Pavla Drázdilová, Jan Martinovic, Vaclav Svaton, Václav SnáselDATESO 2013

TL;DR

This paper introduces a novel method to track the evolution of co-author communities within the DBLP bibliography dataset, using terms from article titles as a proxy for shared research interests. By analyzing these term-defined communities over time, the authors provide a new perspective on how research fields emerge, grow, and stabilize, offering valuable insights for researchers to navigate their professional landscape.

What Problem Does This Paper Address?

Understanding how research communities form and evolve is crucial for researchers to identify emerging trends, stable areas of interest, and promising collaborations. While social network analysis has explored community detection, tracking the dynamic evolution of these communities, especially in the context of academic research, remains a challenge. Existing methods often focus on static snapshots or broad network structures. This paper aims to fill this gap by specifically analyzing the temporal dynamics of co-author communities, using the DBLP dataset and defining communities based on shared keywords from article titles. This approach allows for a more granular understanding of how specific research interests coalesce and change over time.

Methods

The core of this paper lies in its methodology for extracting and analyzing evolving co-author communities. The process begins with the DBLP dataset, from which article titles are processed to extract terms. These terms are then used to define "communities" of authors who have published articles with similar keywords in a given year.

The key steps include:

  • Term Extraction and Lemmatization: Terms from article titles are extracted and lemmatized (using Porter's algorithm) to capture variations of words.
  • Subgraph Generation: For each year, a subgraph is created representing authors and their co-authorship links for articles associated with a specific set of terms.
  • Community Extraction: Connected components within these yearly subgraphs are identified. These components represent the co-author communities for that year and term.
  • Community Evolution Tracking: A crucial algorithm is presented to track the evolution of these communities across consecutive years. It identifies the "follower" component in the next time step that is most similar to the current component, based on shared authors. This allows for the construction of a time-series of communities.
  • Dynamic Metrics: The paper utilizes dynamic metrics, specifically auto-correlation ($C(t)$) and stationarity ($\zeta$), to quantify community evolution. Auto-correlation measures the overlap between a community at time $t$ and $t+\tau$, indicating how stable its membership is. Stationarity, derived from auto-correlation, quantifies the overall stability of a community over its lifespan.

Datasets

The primary dataset used in this research is the DBLP Computer Science Bibliography dataset. The authors processed a significant portion of this dataset, indexing:

  • 2,055,469 articles
  • 1,182,363 authors
  • 308,933 terms from article titles

This dataset was downloaded in October 2012. The paper does not introduce new benchmark datasets but rather leverages the extensive and well-established DBLP data for its analysis.

Key Insights and Ideas

  • Terms as Proxies for Research Interests: The paper demonstrates that terms extracted from article titles can effectively serve as a proxy for identifying shared research interests and thus defining co-author communities. This bypasses the need for explicit community labels or more complex semantic analysis.
  • Dynamic Evolution of Research Fields: The analysis reveals distinct patterns in the evolution of research communities. Some communities, like "social network," show periods of stability followed by growth, while others, like "dynamic social network" and "social network analysis," exhibit more continuous growth.
  • Community Stability vs. Dynamism: The metrics of auto-correlation and stationarity provide a quantitative way to distinguish between stable, long-standing research areas and more dynamic, rapidly evolving ones. For instance, "social network" and "social network analysis" communities are noted as relatively young and thus have higher stationarity values.
  • Generic vs. Specific Terms: The study highlights how generic terms ("social network") can experience fluctuating interest, while more specific terms ("dynamic social network," "social network analysis") might show more consistent growth, suggesting a refinement of research focus over time.

Key Results

The paper presents several key findings based on the analysis of selected terms:

  • Community Lifespans: Table 1 shows the "count of time steps" for various communities, indicating their duration of activity. For example, "elearning teach blackboard" and "elearning teach moodle" communities have the longest observed durations (43 years), spanning from 1971 to 2013.
  • Visualizing Community Growth: Figures 1 and 2 visually represent the number of authors in selected communities over time. These figures clearly illustrate the growth and fluctuations in community size for terms like "social network," "elearning," and "mysql."
  • Auto-correlation Patterns: Figure 3 shows auto-correlation values for different communities. Higher auto-correlation suggests more stable co-authorship over time. For instance, the "elearning" community shows a significant auto-correlation peak from 2011 to 2012, indicating a period of steady interest.
  • Stationarity Values: Figure 4 presents stationarity values, where higher values indicate more static communities. Communities like "social network," "oracle," and "social network analysis" exhibit high stationarity, suggesting a consistent and stable interest among their members.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the DBLP dataset has a limitation: it provides article titles but not abstracts or index terms, which could offer richer semantic information.
  • The algorithm for tracking community evolution relies on identifying the "most similar" component in the next time step. This heuristic might not perfectly capture all nuances of community evolution, especially in cases of significant fragmentation or merging.
  • The analysis of auto-correlation for 2012 could only be calculated at the end of 2013 due to data availability.
  • Future Work:
    • The authors propose enriching their experiments by modifying the community tracking algorithm (Step III in Section 5.1).
  • Incorporating individual user conditions into this step could provide a different perspective on component selection.
  • Analyzing the evolution of communities centered around individual users could offer deeper insights into user characteristics and individual research trajectories.
  • The presented methods are suggested to be applicable to other types of networks and communities beyond academic co-authorship.

Summary updated

BibTeX
@inproceedings{babskova2013evolution,
  author = {Alisa Babskova and Pavla Drázdilová and Jan Martinovic and Vaclav Svaton and Václav Snásel},
  title = {{Evolution of Co-Authors Communities Formed by Terms on DBLP}},
  booktitle = {DATESO},
  year = {2013},
}
Figure