Communication Dans Un Congrès Année : 2025

Integrating Link Prediction and Isolation Forest for Backbone Extraction

Résumé

Backbone extraction simplifies complex networks by preserving essential connections while reducing complexity. Traditional methods rely on fixed thresholds, making them sensitive to parameter choices. Indeed, rare but structurally significant links are often discarded. To address this challenge, we propose a framework that integrates link prediction with anomaly detection. The method assigns scores to links using a similarity-based link prediction function. It then applies Isolation Forest to identify structurally significant links. Unlike conventional approaches, it retains high and low-scoring links, preserving strong connectivity patterns and rare but meaningful interactions. We validate the framework using two link prediction functions: Preferential Attachment and Local Path Index. Three experiments assess its effectiveness. First, we illustrate its ability to preserve central and peripheral links using Zachary’s Karate Club network. Second, we compare it with traditional backbone extraction methods, including the Disparity Filter and the High Salience Skeleton, by evaluating edge and node retention, connectivity, reachability, transitivity, and clustering coefficient deviations. Results show that the Local Path Index backbone best maintains network connectivity and clustering, while the Preferential Attachment backbone provides a more condensed structure. The High Salience Skeleton causes excessive fragmentation, and the Disparity Filter removes too many links, leading to information loss. Third, we analyze how different methods alter edge weight and node degree distributions. The Local Path Index backbone best preserves the original degree distribution, while the Preferential Attachment backbone minimizes edge-weight distortions. The Disparity Filter heavily alters weight distributions, and the High Salience Skeleton significantly modifies degree distributions.The proposed framework balances connectivity and simplification, making it a flexible alternative to traditional backbone extraction methods.

Fichier non déposé

Dates et versions

hal-05291450 , version 1 (01-10-2025)

Identifiants

Citer

Ali Yassin, Hocine Cherifi, Hamida Seba, Olivier Togni. Integrating Link Prediction and Isolation Forest for Backbone Extraction. Modelling and Mining Networks, Jun 2025, Vilnius, Lithuania. pp.189-204, ⟨10.1007/978-3-031-92898-7_13⟩. ⟨hal-05291450⟩
107 Consultations
0 Téléchargements

Altmetric

Partager

  • More