Introduction
Ranking objects within ontologies presents a unique challenge in the Semantic Web. Traditional approaches typically diverge into two distinct methodologies: link-based ranking and preference-based ranking. Link-based systems, such as Google’s PageRank, provide an objective measure of "importance" or "authority" based on the structural topology of the graph. However, these methods often ignore specific user intent. Conversely, preference-based systems allow for subjective criteria but fail to account for the global importance or quality of the objects.
This paper proposes a unified framework that merges conditional user preferences with link-based analysis, while leveraging probabilistic description logics to handle the inherent incompleteness of real-world ontological data.
Methodology
The proposed framework integrates three core components: conditional preferences, hybrid ranking strategies, and probabilistic reasoning.
Variable-Strength Conditional Preferences
The system allows users to define preferences within specific contexts. These are expressed as conditional statements with associated strengths (e.g., "In the context of academic publications, international venues are preferred over national ones with a strength of 70"). This allows for a nuanced representation of user intent that varies across different domains of the ontology.
Hybrid Integration Strategies
The authors propose two distinct methods for combining user preferences with the PageRank algorithm:
- Preference-Dominated Ranking: In this approach, objects are first categorized into tiers based on their preference scores. PageRank is then utilized as a secondary sorting mechanism to break ties within each tier. This ensures that user-defined criteria are the primary driver of the results.
- PageRank-Dominated (Personalized) Ranking: This method integrates preference scores directly into the PageRank formula. By substituting the standard uniform "teleportation" vector ($E(u)$) with a vector biased toward the user's preferred objects, the "random surfer" is more likely to jump to nodes that align with user interests. This effectively personalizes the global importance metric.
Probabilistic Description Logics
To address the "Open World Assumption"—where missing information does not necessarily imply the absence of a property—the framework employs a combination of DL-Lite (a computationally efficient logic) and Bayesian networks. This allows the system to calculate the probability of an object belonging to a specific class. By computing the "expected rank" based on these probabilities, the system can effectively rank objects even when their metadata is incomplete.
Theoretical Application and Insights
The framework is illustrated through a literature search scenario, ranking publications based on metadata such as titles and venue types. A key insight of this work is the personalization of the "random surfer" model; by biasing the teleportation vector, the algorithm reflects a user's specific search context without losing the structural benefits of link analysis.
Furthermore, the use of expected concept membership provides a robust mechanism for ranking in the face of uncertainty. The system naturally categorizes results into tiers (e.g., International > National > Non-conference), providing a structured way to navigate search results beyond a simple flat list.
Conclusion and Future Work
While the paper provides a rigorous theoretical foundation for hybrid ranking, several areas for future research remain. The current work lacks large-scale empirical validation or user studies to quantify improvements in search precision or user satisfaction. Additionally, the computational overhead of combining Bayesian networks with description logics may pose challenges for very large ontologies. Future efforts should also focus on the systematic tuning of preference weights and their mathematical interaction with the PageRank damping factor ($c$).
Summary updated