A research team led by Prof. QUI Anqi, Director of Mental Health Research Centre (MHRC), Associate Dean of Graduate School, Professor of Department of Health Technology and Informatics, has developed a novel heterogeneous graph attention network namely “Hodge-Laplacian Heterogeneous Graph Attention Network” (HL-HGAT), which is set to revolutionise the modelling of complex relationships in graph-structured data. This innovation is poised to overcome the current limitations in fields such as logistics, computer vision, chemistry and neuroscience.
Traditional graph neural networks (GNNs) mostly look at simple “A-to-B” type connections, which makes it hard for them to understand group interactions. Prof. Qiu’s new HL-HGAT model goes beyond this by interpreting graphs as higher-dimensional shapes (called simplicial complexes), so it can capture relationships among nodes, edges, triangles and higher-order structures. Central to HL-HGAT is the Hodge-Laplacian operator, which facilitates the modelling and propagation of signals beyond simple pairwise relationships, offering a richer understanding of complex data. A key innovation of HL-HGAT is its capacity to model dynamic, time-varying graphs, using HL filtering, adaptive attention and heterogeneous signal decomposition to reveal evolving patterns that static GNNs may overlook.
The model has demonstrated versatility across a range of applications, including logistics (optimising delivery routes), computer vision (improving image classification), chemistry (predicting molecular properties), and neuroscience (analysing brain imaging data). Notably, HL-HGAT can detect subtle neural changes in conditions like depression and Alzheimer’s disease, outperforming traditional methods and enabling earlier diagnosis and intervention in clinical settings.
This innovative HL-HGAT model not only achieves outstanding results in addressing complex graph-based tasks in both scientific and industrial domains, but also represents a significant advancement in GNN technology. The research, detailed in a paper titled “HL-HGAT: Heterogeneous Graph Attention Network via Hodge-Laplacian Operator”, has been published in IEEE Transactions on Pattern Analysis and Machine Intelligence.
Press release: https://polyu.me/49OzR1u
Online coverage:
Mirage - https://polyu.me/4k68Cmp
Hong Kong Economic Journal - https://polyu.me/49XexFD (subscription required)
Wen Wei Po - https://polyu.me/4a1dapz
Hong Kong Commercial Daily - https://polyu.me/4afQ3c5
Bastille Post - https://polyu.me/49OAzfa
| Research Units | Mental Health Research Centre |
|---|
