| 作成者 |
|
|
|
| 本文言語 |
|
| 会議情報 |
|
|
|
| アクセス権 |
|
| 権利関係 |
|
|
|
| 関連DOI |
|
| 関連URI |
|
| 概要 |
With the rise of the Internet, fraudulent activities like cherry-picked reviews, fake accounts, and credit card fraud have become increasingly sophisticated. Graph Neural Networks (GNNs) have been wid...ely used for fraud detection, but common models (e.g., GCN, GAT) struggle due to the low homophily problem in graphs. To address this, two main types of fraud detectionspecific GNN models have been proposed: (1) filtering methods that create highly homophilic environments by selecting specific neighbors and (2) edge classification methods that partition graphs into homophilic and heterophilic subgraphs. However, filtering may lead to information loss, and edge classification struggles with accurate partitioning. This study proposes a novel method that dynamically partitions the graph into two subgraphs based on the distribution of neighboring nodes. Unlike prior approaches, our method adapts to the graph structure, improving the quality of embedded representations for classification. We evaluate our model using fraud detection datasets from Yelp and Amazon. Experimental results show that our approach outperforms existing methods in classification performance, achieving superior AUC and AP scores.続きを見る
|
Mendeley出力
