中国科学院自动化研究所机构知识库

In the era of big data, large-scale data can be very effective in improving model performance. However, in the real world, high-quality data is usually difficult to acquire due to privacy or cost. Especially when it comes to credit card fraud, the fraud samples are quite rare. Detecting card fraud with few samples is a meaningful task. Graph neural network (GNN) is a good way to deal with few samples because an advantage of GNN is that information can be disseminated through connections between nodes. However, the data structure of credit cards cannot be applied by the GNN-based method directly. In this paper, we proposed a GNN-based few-shot learning method which can detect credit card fraud with few samples effectively. We constructed a learnable parametric adjacency matrix method relying on the similarity of features to pass messages and utilized the GCN layer to extract node features. We compared our method with classical machine learning algorithms and other graph neural networks on the real-world data set. Our experimental results show that our proposed model can perform better extremely with fewer training samples than baselines.