Molecular Informatics, Journal Year: 2025, Volume and Issue: 44(3)
Published: March 1, 2025
Abstract Within a recent decade, graph neural network (GNN) has emerged as powerful architecture for various graph‐structured data modelling and task‐driven representation learning problems. Recent studies have highlighted the remarkable capabilities of GNNs in handling complex tasks, achieving state‐of‐the‐art results node/graph classification, regression, generation. However, most traditional GNN‐based architectures like GCN GraphSAGE still faced several challenges related to capability preserving multi‐scaled topological structures. These models primarily focus on capturing local neighborhood information, often failing retain global structural features essential graph‐level classification tasks. Furthermore, their expressiveness is limited when structures molecular datasets. To overcome these limitations, this paper, we proposed novel which an integration between neuro‐ f uzzy t o p ological g raph approach, naming as: FTPG. Specifically, within our FTPG model, introduce approach property prediction by integrating with advanced components. The employs separate modules effectively capture both graph‐based well features. Moreover, further address feature uncertainty global‐view representation, multi‐layered neuro‐fuzzy incorporated model enhance robustness learned embeddings. This combinatorial can assist leverage strengths multi‐view multi‐modal learning, enabling deliver superior performance Extensive experiments real‐world/benchmark datasets demonstrate effectiveness model. It consistently outperforms baselines categorized different approaches, including canonical proximity message passing based, transformer‐based, topology‐driven approaches.
Language: Английский