PAV: A novel model for ranking heterogeneous objects in bibliographic information networks

Zhi-Hong Deng, Bo-Yan Lai, Zhonghui Wang, Guo-Dong FangExpert Systems with Applications 2012 · DOI

TL;DR

This paper introduces PAV, a novel model for ranking diverse objects (authors, papers, venues) within bibliographic information networks. By representing these networks as a weighted directed graph and employing a regular Markov chain, PAV effectively captures complex relationships to provide a unified and more accurate ranking system than existing methods.

What Problem Does This Paper Address?

Bibliographic information networks, like those found in ACM Digital Library or IEEE Xplore, are rich repositories of interconnected data about authors, papers, and publication venues. While these networks are extensively studied for management and data mining tasks, there's a significant gap in research concerning how to effectively rank the objects within them. Existing ranking methods, such as the impact factor for journals or the h-index for authors, are typically designed for homogeneous data and struggle to account for the intricate, heterogeneous relationships present in bibliographic networks (e.g., an author's reputation being influenced by their co-authors, the venues they publish in, and the papers they cite). This limitation hinders applications like improved search, recommendation systems, and a deeper understanding of academic influence.

Methods

The core of the PAV model lies in its representation of a bibliographic information network as a weighted directed graph.

  • Graph Construction:
    • Nodes: Represent authors, papers, and venues.
  • Edges: Capture relationships like "author writes paper," "paper cites paper," "paper published by venue," and "author associated with venue." Crucially, PAV creates bidirectional edges for most relationships to reflect mutual influence.
  • Weights: Edges are assigned weights that quantify the contribution of one object to another's importance. These weights are defined based on factors like an author's position in a paper's author list, co-authorship patterns, and the number of papers shared between an author and a venue.
  • Ranking Mechanism:
    • The problem of ranking is transformed into estimating the probability distribution of accessing nodes in the graph.
  • This is achieved by treating the PAV graph as a regular Markov chain. A random walker navigates the graph, with the probability of moving between nodes proportional to the edge weights.
  • The transition probability matrix (TPM) is constructed based on these edge weights and a small probability of random "jumping" to any node, ensuring connectivity.
  • The stationary distribution of this Markov chain, calculated iteratively, represents the importance or reputation of each object. The process converges to a stable distribution, meaning the ranking is independent of the initial starting point.

Datasets

  • ACM Digital Library Portal: A real-world dataset crawled from ACM Digital Library, containing metadata for papers published from 1950 to 2008. This dataset includes 196,044 authors, 162,256 papers, and 636 conferences. The authors specifically note that DBLP was not used because it lacks citation information, which is crucial for their model.

Key Insights and Ideas

  • Unified Ranking for Heterogeneity: The most significant insight is the development of a single framework (PAV) capable of ranking diverse entities (authors, papers, venues) simultaneously, overcoming the limitations of single-purpose ranking metrics.
  • Bidirectional Influence: By using bidirectional edges, PAV acknowledges that influence flows in multiple directions. For instance, a venue's reputation is influenced by the papers it publishes, and conversely, the papers published in a venue contribute to its reputation.
  • Markov Chain for Dynamic Importance: The application of a regular Markov chain elegantly models the dynamic nature of influence and reputation within the network, where importance is a probabilistic outcome of navigation.
  • Weighting Scheme Captures Nuance: The carefully designed weighting scheme for edges, considering factors like author order and co-authorship, allows the model to capture subtle differences in relationships that simpler metrics might miss.

Key Results

The paper demonstrates the effectiveness of the PAV model through experimental evaluation on the ACM Digital Library dataset.

  • Conference Ranking: PAV's conference rankings showed strong consistency with established systems like CSRank, CiteSeer Rating, and ArnetMiner Rank across different areas (Applications & Media, Databases, Hardware & Architecture). For example, SIGGRAPH was ranked highly by PAV and other systems, reflecting its recognized importance.
  • Author Ranking: PAV's author rankings differed significantly from ArnetMiner's H-index rankings. The paper attributes this to PAV's ability to incorporate a broader range of influences (co-authors, venues, papers) beyond just the author's publication count. For instance, an author with fewer papers but more publications in top venues like SIGMOD received a higher PAV rank.
  • Paper Ranking: Top-ranked papers in the database and information retrieval areas were consistently published in prestigious venues (SIGMOD, SIGIR) and had high citation counts, aligning with expert intuition.

Limitations and Future Work

  • Date Impact: The authors acknowledge that their current model does not explicitly consider the time factor (i.e., older papers might naturally have more citations). They suggest extending the model with time-oriented decay factors or by comparing papers published within similar timeframes.
  • Data Completeness: While the ACM DL dataset is substantial, the effectiveness of PAV, like any network-based model, is dependent on the completeness and accuracy of the underlying bibliographic data.
  • Future Work: The authors propose extending PAV to other mining tasks such as clustering and classification. They also plan to apply and further study the model's performance on other bibliographic databases.

Summary updated

BibTeX
@article{deng2012pav,
  author = {Zhi-Hong Deng and Bo-Yan Lai and Zhonghui Wang and Guo-Dong Fang},
  title = {{PAV: A novel model for ranking heterogeneous objects in bibliographic information networks}},
  journal = {Expert Systems with Applications},
  year = {2012},
  doi = {10.1016/j.eswa.2012.02.175},
}
Figure