Abstract
Link prediction in large-scale social networks faces a significant trade-off between semantic richness and computational scalability. This paper evaluates the "Assertion Role"—a strategy for selecting a minimal subset of semantically relevant individuals to ground a probabilistic description logic model. Using the $CR\mathcal{ALC}$ framework on real-world scientific curricula data, we demonstrate that hybrid selection metrics combining graph topology with semantic filtering can achieve over 99% prediction accuracy while reducing the computational load by a factor of 50 compared to purely structural approaches.
Introduction
Link prediction—the task of forecasting future connections between nodes in a network—is a cornerstone of network science. While traditional methods rely on graph topology (e.g., common neighbors), integrating semantic information (e.g., shared research interests) significantly improves accuracy. However, incorporating this data through probabilistic models often introduces exponential computational complexity. This paper addresses the question: Which individuals actually matter when predicting a specific link? By filtering out irrelevant data (assertions), we can maintain model scalability without losing predictive power.
Methodology
The researchers utilize the $CR\mathcal{ALC}$ (Credal $\mathcal{ALC}$) framework, a probabilistic description logic that allows for ontologies with uncertainty. For instance, one might define a 91% probability that researchers sharing a publication are collaborators.
The core of the approach is the "Assertion Role," which defines the subset of individuals ($\mathcal{A}(a, b)$) used to ground the probabilistic model for a target pair $(a, b)$. Four selection strategies were tested:
- Structural (Simple): Immediate neighbors of the target nodes.
- Structural (Path-based): Individuals on paths between nodes plus their neighbors.
- Hybrid (Semantic Filter of 1): Neighbors filtered for the top 10 most semantically similar individuals.
- Hybrid (Semantic Filter of 2): Path-based group filtered for the top 10 most semantically similar.
Datasets
The study used data from the Lattes platform, a repository of Brazilian scientific curricula, focusing on 8,000 researchers and 95,000 co-authorship links. Semantic similarity was calculated using keyword matches from publication titles. A balanced dataset of 2,000 links was used for evaluation via a Logistic Regression classifier.
Results and Discussion
The study found that including hundreds of distantly related individuals often adds noise rather than predictive value. The most striking insight is the efficiency of semantic filtering.
| Feature / Metric | Accuracy | Avg. # of Individuals Selected |
|---|---|---|
| CRALC + Metric i (Structural) | 99.93% | 501 |
| CRALC + Metric iii (Hybrid/Filtered) | 99.88% | 10 |
| CRALC + Metric vi (Random) | 71.14% | 10 |
Metric iii achieved virtually the same accuracy as the much larger structural model ($>99.8\%$), but required processing only 10 individuals instead of 501. This represents a 50x improvement in efficiency with negligible loss in precision.
Conclusion
Semantic filtering is a powerful tool for pruning the search space in probabilistic link prediction. Combining graph structure with semantic logic consistently outperforms random selection or purely path-based methods. Future work will explore more complex ontologies and advanced NLP techniques to further refine the selection process.
Summary updated