Associations Dynamic Evolution: Evolving Graph Transformer

Cheng Wang ()
Additional contact information
Cheng Wang: Tongji University

Chapter Chapter 8 in Anti-Fraud Engineering for Digital Finance, 2023, pp 189-207 from Springer

Abstract: Abstract In online lending services, fraud prediction is an especially critical step to control loss risk and improve processing efficiency. It is definitely challenging that such predictions need to detect evolving and increasingly impalpable fraud patterns. The technical difficulty mainly stems from one factor: evolution of fraud patterns. As a widely recognized method currently, GNNs has attracted much attention from researchers. According to the requirements of the task scenario, graphs can be divided into four categories from two perspectives: static-homogeneous graph, static-heterogeneous graph, dynamic-homogeneous graph, and dynamic-heterogeneous graph. The GNNs on the dynamic-heterogeneous graph is undoubtedly the methods with the highest level and the most generalization. From the perspective of heterogeneity, the past methods mainly used meta-path (the disadvantage is that additional expert knowledge is required), and newer works (such as HGT) abandon meta-path and use meta-relation instead, and set up multiple sets of projections The type of matrix modeling edge. From the perspective of dynamics, the past methods mainly used the sequential combination of GNN+RNN (such as TGCN), but this method is difficult to learn irregular behavior. A new approach is not to train the embedding of the GNN output, but to use the RNN to train the parameters of the GNN (making the model itself have higher generalization), while also reducing the number of parameters that need to be learned. We propose Evolving Graph Transformer (EGT), which has three main improvements. It does not rely on expert knowledge by using the idea of meta-relation to directly model the node pair. In order to better model the diversity of node and edge types in heterogeneous graphs, we set up each type-specific projection matrix so that their distribution can be displayed more realistically. Compared with using RNN to learn the node representation of GNN output, we use RNN to directly evolve the parameters of GNN. This approach effectively performs model adaptation, which focuses on the model itself rather than the node embeddings.

Date: 2023
References: Add references at CitEc
Citations:

There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:spr:sprchp:978-981-99-5257-1_8

Ordering information: This item can be ordered from
http://www.springer.com/9789819952571

DOI: 10.1007/978-981-99-5257-1_8

Access Statistics for this chapter

More chapters in Springer Books from Springer
Bibliographic data for series maintained by Sonal Shukla () and Springer Nature Abstracting and Indexing ().