.png){kind=icon}
Return
3PIGENet:A multi-pollutant identification algorithm using multi-source spectroscopy for label-missing scenarios
.png){kind=icon}
OA
AI
DOI：10.1016/j.wroa.2025.100359.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Real-world water samples often face issues such as multi-source pollution mixing and missing labels, posing significant challenges to the accurate identification of pollutants. In this context, the key to pollutant identification lies in effectively representing the spectral information of pollutants and separating features of different categories through semi-supervised learning strategies. Therefore, this study proposes a semi-supervised prototype contrast loss Graph Convolutional Network (GCN) model based on multi-source spectroscopy for the identification of water pollutants. The method first constructs a spectral library of standard pollution samples and extracts topological features of samples using the characteristic peak relationship function of different pollution sources. By integrating GCN, the model can calculate prototype vectors under different pollution conditions and perform semi-supervised learning by combining prototypical contrastive learning with cross-entropy loss to address sample pollution representation in label-missing scenarios. Finally, the prediction of pollutants is made by calculating the cosine similarity between the learned pollution prototype vectors and the pollution sample representation vectors. Compared to traditional spectral feature extraction methods, topological features can better reveal the relationships between characteristic peaks in different spectra. Graph prototype contrastive learning aids in more robustly capturing inter-class discrimination information and feature consistency. Experimental results indicate that in water pollutant identification tasks, especially in semi-supervised learning tasks, proposed 3PIGENet model achieves an accuracy of 0.915 and a macro-F1 score of 0.683 in semi-supervised learning, which significantly outperforms existing SOTA models.
Keywords:
Graph convolutional network (GCN)
Semi-supervised learning
Prototype Contrast Learning
UV–Vis
EEM
Journal
IF:
12.4
Papers:
3.1W
Citations:
15.7W
Organization
Cited Papers
Cross-Social-Network User Identification Based on Bidirectional GCN and MNF-UI Models.png)
Electronics
OAAI
ElectronicsIF0
OAAIShareSaveShareSaveA review of graph neural networks: concepts, architectures, techniques, challenges, datasets, applications, and future directionsJOURNAL OF BIG DATAOAAI
JOURNAL OF BIG DATAIF6.4
OAAIShareSaveAntibiotic pollution of the Yellow River in China and its relationship with dissolved organic matter: Distribution and Source identificationWATER RESEARCH
WATER RESEARCHIF12.4
PREAI
ShareSaveShareSaveShareSave
View more