Institutional Repository of Vilnius University
Applying graph convolutional networks for community detection: quantifying node influence /
Abstract
dc:descriptionThis thesis extends a Graph Convolutional Network (GCN) implementation for community detection with three additions, all aimed at figuring out which nodes the model's decisions actually depend on. The first is a switch to structural features, which lets the same trained model run on graphs of any size. The second is a leave-one-out procedure that gives each node an influence score based on how much its removal disturbs the model's clustering. The third is a frozen-inference perturbation experiment, where the same trained model is evaluated on a graph extended with a new community, and Spearman's rank correlation compares the influence rankings before and after. Experiments on Zachary's Karate Club and a custom 80-node graph produced three findings: The single-scale variant beat the multi-scale one on both graphs, since the structural features already carry multi-hop information and adding A^2 and A^3 on top just pushes the model into over-smoothing. The GCN implementation's notion of which nodes are important does not line up with classical centrality - peripheral and boundary nodes dominated the rankings rather than the high-degree hubs. Influence stability under perturbation tracks model quality: the one configuration where the model worked properly at baseline was also the one where its rankings survived perturbation cleanly, while the configurations that failed at baseline ended up with rankings that were either uncorrelated or actually inverted once the graph was extended.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gainulin, Nikita,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308337267&prefLang=en_US