Isolation Forest Meets Link Prediction: A Novel Framework for Backbone Extraction
Résumé
Complex networks appear across various disciplines, including social sciences, biology, finance, and transportation. Their complexity often obscures critical structural and functional relationships, making analysis difficult. Backbone extraction simplifies networks by preserving key connections while reducing complexity. Traditional backbone extraction methods apply fixed thresholds on computed scores, making them highly sensitive to parameter choices. In a previous work we introduced similarity-based and embedding-based link prediction techniques for backbone extraction. However, these methods depend on predefined thresholds. They preserve either the top or bottom fraction of scores and discard the remaining fraction, which may contain structurally significant interactions. To overcome these limitations, this study presents a novel backbone extraction framework integrating link prediction with anomaly detection.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |