Szczegóły publikacji
Opis bibliograficzny
Harnessing Graph Neural Networks: a comparative study of transformers and autoencoders in link prediction for high-energy physics / Wojciech GOMUŁKA, Maciej KRZYWDA, Piotr A. KOWALSKI, Tomasz SZUMLAK, Tomasz BOŁD, Szymon ŁUKASIK, Amir H. Gandomi // W: Advances in artificial intelligence research : proceedings of the 6th Polish conference on Artificial Intelligence, PP-RAI 2025, Katowice, Poland, 7–9 April, 2025 / eds. Beata Zielosko, [et al.]. — Cham : Springer Nature Swizterland, cop. 2026. — ( Lecture Notes in Networks and Systems ; ISSN 2367-3370 ; vol. 1599 ). — ISBN: 978-3-032-04196-8; e-ISBN: 978-3-032-04197-5. — S. 219–228. — Bibliogr., Abstr. — Publikacja dostępna online od: 2026-01-02. — W. Gomułka - dod. afiliacja: Center of Excellence in Artifical Intelligence, AGH University of Krakow. – P. A. Kowalski, Sz. Łukasik - dod. afiliacja: Polish Academy of Sciences, Systems Research Institute, Warsaw
Autorzy (7)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 165509 |
|---|---|
| Data dodania do BaDAP | 2026-02-19 |
| DOI | 10.1007/978-3-032-04197-5_17 |
| Rok publikacji | 2026 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Springer |
| Czasopismo/seria | Lecture Notes in Networks and Systems |
Abstract
Graph Neural Networks (GNNs) are a trending topic in track reconstruction in high-energy physics (HEP). In recent years, GNNs have proven to be an interesting alternative to traditional approaches. In this paper, we would like to present the continuation of our previous analysis devoted to edge prediction in HEP tracking data. Therefore, in this case study, the research is extended through training models with multiple collision events, providing comparisons of the results obtained for different data volumes of the events (so-called pile-up) and testing other neural networks. In the very case study, we put emphasis on GNN analouges of autoencoders and transformers. Experiments were conducted on simulated proton-proton collision events generated with the ACTS toolkit, analyzing datasets with pile-up values of 10 and 200. Several architectures, including Graph Transformer, Graph Autoencoder, Variational Graph Autoencoder, Adversarially Regularized Graph Autoencoder, and Adversarially Regularized Variational Graph Autoencoder, were evaluated in terms of precision, recall, specificity, AUC-ROC, and AUC-PR metrics. Training was performed multiple times for stability assessment, and the results suggest that non-variational models, particularly Graph Transformer, ARGA, and GAE, achieved superior performance in distinguishing real and false edges. The impact of dataset size was also evident, with models trained on higher pile-up data exhibiting improved accuracy. These findings highlight the potential of transformer-based architectures for large-scale HEP tracking applications while emphasizing the necessity for further scalability studies and model optimizations in realistic reconstruction pipelines.
