On the Geometric Coherence of Global Aggregation in Federated GNN

Chethana Prasad Kabgere, Shylaja S S · 2026

Read On the Geometric Coherence of Global Aggregation in Federated GNN by Chethana Prasad Kabgere, Shylaja S S (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15510.txt.

Description

Kabgere and Shylaja point to a mismatch at the heart of federated learning on graphs. Standard federated methods average model weights as if they lived in one shared flat space, but graph neural network weights act as message-passing operators whose sense hangs on graph shape. When client graphs differ in shape and in data mix, each local update perturbs a distinct operator, so plain averaging blends directions that do not fit together.

The result, they argue, is a global model that may converge and score well yet lose feel for graph structure, a loss that loss and accuracy scores miss. Their fix is GGRS, a server-side scheme that works on a data-free proxy of operator shifts and enforces fit through alignment, subspace match, and sensitivity control. The aim is to keep message passing coherent while clients keep their data local.

Catalog record

CREATORChethana Prasad Kabgere, Shylaja S S
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-15510

Browse the full library · About the Tower of Babel