Intelligent Systems.
Theory and Applications

(Intellektual'nye Sistemy. Teoriya i Prilozheniya)

The influence of adjacency matrix construction on the generalization ability of graph neural networks in wearable sensor signal analysis

Abstract

The paper tests how the choice of the adjacency matrix affects the macro-F1 of a graph neural network. On UCI HAR, over \(20\) paired runs, a channel-metadata graph is compared with its equal-degree complement, a graph without edges, and hybrid fine-tuning. The mean difference from the complement is \(1.5\)–\(1.9\) percentage points and is significant after the Holm correction at every training fraction. The difference from the graph without edges is significant at \(5\) and \(10\%\). Hybrid fine-tuning is significantly worse than the fixed matrix at \(10\) and \(100\%\). The spread of the difference within one subsample exceeds the spread between subsamples. A sweep of initial graphs on the full training set shows no dip.

Keywords: graph neural networks, adjacency matrix, multivariate time series, generalization ability, activity recognition.

BibTeX
@article{IS-Sidorov-Maysuradze2026,
  author  = {Sidorov, Leonid Stanislavovich and Maysuradze, Archil Iverievich},
  title   = {{The influence of adjacency matrix construction on the generalization ability of graph neural networks in wearable sensor signal analysis}},
  journal = {Intelligent Systems. Theory and Applications},
  year    = {2026},
  volume  = {30},
  number  = {4},
  pages   = {413--416},
}
AMSBIB
\Bibitem{IS-Sidorov-Maysuradze2026}
\by L.\,S.~Sidorov, A.\,I.~Maysuradze
\paper The influence of adjacency matrix construction on the generalization ability of graph neural networks in wearable sensor signal analysis
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 413--416
\lang In Russian
Published under Creative Commons Attribution 4.0 International (CC BY 4.0)

← Back to issue

× Issue cover