Fig 1.
This is an example of a telecom graph network structure.
Black users are fraud users, white users are normal users, and the line represents the calling relationship between users.
Fig 2.
An illustration is provided here to show the details of the proposed GDFGAT model. The model is divided into four modules.
(1) Metadata module. (2) Feature extraction module, which extracts features from two data set dimensions. (3) The main model module consists of the GRU model and the GAT model. (4) Classification module, which classifies each node.
Fig 3.
Determination process of feature difference weight.
Table 1.
Sichuan telecom fraud dataset static analysis.
Table 2.
CDR dataset features information.
Table 3.
Features of the Sichuan dataset after extraction.
Fig 4.
The x-axis on the left side of
Fig 4 is the proportion of sample sampling, and the y-axis is the proportion of joint calls made by fraudulent users and non-fraudulent users. The x-axis on the right side of Fig 4 is the proportion of sample sampling, and the y-axis is the proportion of fraudulent users and non-fraudulent users using second-order neighbors or fraudulent users.
Fig 5.
The x-axis represents time, from August 2019 to March 2020, and the y-axis represents the total size of feature statistics of fraudsters or non-fraudsters each month.
Table 4.
Performance comparison of GDFGAT and baseline on SiChuan dataset.
Table 5.
Static analysis of Yelp and Amazon data sets.
Table 6.
Performance comparison of GDFGAT and baseline on Amazon and Yelp dataset.
Fig 6.
Ablation experiment diagram on SiChuan, Amazon and Yelp data sets.
The x-axis represents each metric and the y-axis represents the corresponding value. The SiChuan dataset has five metrics, and the Yelp and Amazon data sets have four metrics.
Fig 7.
Performance of GDFGAT with different embedding sizes, layers, and multi-head on the Sichuan dataset.
In Fig(a), the x-axis represents the embedding size, and the y-axis represents the corresponding value. In Fig(b), the x-axis represents the number of graph neural layers, ranging from 1 to 5, and the y-axis represents the corresponding value. In Fig(c), the x-axis represents the number of heads of the graph neural network, ranging from 1 to 6, and the y-axis represents the corresponding value.