Institution profile

Istituto per le Applicazioni del Calcolo 'Mauro Picone', Consiglio Nazionale delle Ricerche

Academic institutioneurope · it
Official website
Research library2linked papers
Opportunities0open roles
Selected work

Representative Papers

Random Hyperbolic Graphs with Arbitrary Mesoscale Structures

Jun 03, 2025

Real-world networks exhibit both geometric properties—such as sparsity, small-worldness, power-law degree distribution, and high clustering—and non-geometric mesoscale structures—e.g., arbitrary community mixing patterns. Classical Random Hyperbolic Graphs (RHGs) model only node similarity and popularity, but their strict triangle inequality constraint prevents accurate representation of non-geometric inter-community connections. To address this, we propose the Random Hyperbolic Block Model (RHBM), the first model to explicitly integrate block structure into hyperbolic geometry under a maximum-entropy framework. RHBM decouples intra- and inter-block similarity, enabling violation of the triangle inequality to capture realistic inter-community links. Experiments demonstrate that RHBM preserves RHG’s macroscopic properties while precisely and controllably reproducing target community structures—outperforming RHG significantly on synthetic benchmarks.

0 citationsRead paper

TROPIC -- Trustworthiness Rating of Online Publishers through online Interactions Calculation

Jan 23, 2025

To address the high cost and poor scalability of news source credibility assessment, this paper proposes the first framework for modeling news outlet credibility based on large-scale social media user interaction behaviors (Twitter/X and Reddit). Methodologically, it integrates graph neural networks with interaction sequence modeling, incorporating collaborative signal mining and an explainable feedback mechanism to enable zero-shot initial evaluation for cold-start news sources and active-learning-driven iterative refinement. An interactive annotation platform is further developed to facilitate human verification. Experiments demonstrate cross-platform rating consistency of 86.4% (vs. expert panel), a 79.2% accuracy for cold-start initial evaluation, and a 3.8× improvement in human annotation efficiency. This work establishes a new paradigm for scalable, interpretable, and minimally human-dependent news credibility assessment.

0 citationsRead paper
Recent publications

Latest Papers

Random Hyperbolic Graphs with Arbitrary Mesoscale Structures

Jun 03, 2025

Real-world networks exhibit both geometric properties—such as sparsity, small-worldness, power-law degree distribution, and high clustering—and non-geometric mesoscale structures—e.g., arbitrary community mixing patterns. Classical Random Hyperbolic Graphs (RHGs) model only node similarity and popularity, but their strict triangle inequality constraint prevents accurate representation of non-geometric inter-community connections. To address this, we propose the Random Hyperbolic Block Model (RHBM), the first model to explicitly integrate block structure into hyperbolic geometry under a maximum-entropy framework. RHBM decouples intra- and inter-block similarity, enabling violation of the triangle inequality to capture realistic inter-community links. Experiments demonstrate that RHBM preserves RHG’s macroscopic properties while precisely and controllably reproducing target community structures—outperforming RHG significantly on synthetic benchmarks.

0 citationsRead paper

TROPIC -- Trustworthiness Rating of Online Publishers through online Interactions Calculation

Jan 23, 2025

To address the high cost and poor scalability of news source credibility assessment, this paper proposes the first framework for modeling news outlet credibility based on large-scale social media user interaction behaviors (Twitter/X and Reddit). Methodologically, it integrates graph neural networks with interaction sequence modeling, incorporating collaborative signal mining and an explainable feedback mechanism to enable zero-shot initial evaluation for cold-start news sources and active-learning-driven iterative refinement. An interactive annotation platform is further developed to facilitate human verification. Experiments demonstrate cross-platform rating consistency of 86.4% (vs. expert panel), a 79.2% accuracy for cold-start initial evaluation, and a 3.8× improvement in human annotation efficiency. This work establishes a new paradigm for scalable, interpretable, and minimally human-dependent news credibility assessment.

0 citationsRead paper