Szczegóły publikacji
Opis bibliograficzny
Graph neural network architecture search via hybrid genetic algorithm with parallel tempering / Maciej KRZYWDA // W: CIKM'25 : proceedings of the 34th ACM international conference on Information and Knowledge Management : November 10 - 14, 2025, Seoul, Republic of Korea / eds. Meeyoung Cha, [et al.]. — New York : Association for Computing Machinery, Inc. (ACM), cop. 2025. — ISBN: 979-8-4007-2040-6. — S. 6793–6796. — Bibliogr. s. 6795, Abstr. — Publikacja dostępna online od: 2025-11-10
Autor
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 164416 |
|---|---|
| Data dodania do BaDAP | 2025-12-10 |
| Tekst źródłowy | URL |
| DOI | 10.1145/3746252.3761661 |
| Rok publikacji | 2025 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Creative Commons | |
| Wydawca | Association for Computing Machinery (ACM) |
| Konferencja | ACM International Conference on Information and Knowledge Management 2025 |
Abstract
In recent years, there has been a surge of interest in harnessing graph neural networks (GNNs) for graph classification tasks. Despite the strong performance of manually designed GNN architectures, their development often relies on time-consuming, expert-driven trial and error, which may overlook promising design opportunities. To address this challenge, we propose a hybrid Genetic Algorithm and Parallel Tempering (GA+PT) framework for automated GNN architecture search. Our method systematically encodes a rich design space-including convolutional layer types (GCN, GraphSAGE, GAT, GIN, SGC), attention heads, hidden-layer widths, dropout rates, weight-initialization schemes, learning rates, and classifier structures-into evolutionary genotypes. A population-based GA explores this space via crossover and mutation, while an inner PT Markov Chain Monte Carlo procedure adaptively refines individual solutions under a temperature schedule to escape local optima. We validate our approach on benchmark graph datasets, demonstrating its ability to discover high-performing architectures that balance classification accuracy, macro-F1 score, and model complexity. The proposed GA+PT hybrid offers a robust, scalable, and resource-aware alternative to manual GNN design and purely gradient-based neural architecture search methods.
