| With the development of information technology,people’s lives are increasingly dependent on the Internet.However,various abnormal events on the Internet often cause losses to people.In order to detect anomalies on data organized in network form,it is necessary to represent the embedding of the network.So far,many static network representation methods have been proposed,which can capture the organizational patterns of the topology and attributes in the network through these models.However,for real networks that constantly change over time in daily life,such static methods cannot learn the evolutionary characteristics of the network.Among the currently proposed methods for dynamic network anomaly detection,most encode the time information and input it into the model,and use the method of recurrent neural networks to learn the evolutionary pattern.Such methods often have two problems,for example,they do not explicitly model events in the network,but directly learn the evolutionary pattern of the network through network snapshots,ignoring the short-term and long-term impact of abnormal events on the network structure.In addition,such methods mostly use both recurrent neural networks and graph neural networks to model the topology and time information of the network respectively,ignoring the mutual influence between the two.To solve these problems,this thesis proposes a relevant algorithm,which makes the following contributions:(1)To address the problem that existing models are not effective in modeling abnormal events on dynamic networks,this thesis proposes a dynamic network anomaly detection algorithm based on Hawkes processes(DGHP).DGHP uses spatio-temporal sampling and Hawkes processes to model abnormal events on dynamic networks,and can more effectively characterize connection events on dynamic networks.Experiments show that DGHP outperforms current state-of-the-art models by more than 1% in dynamic network link prediction tasks.(2)To address the problem that current models cannot effectively learn both the topology and temporal information of dynamic networks,this thesis proposes a new dynamic network anomaly detection model,Dyn Trans,by migrating the structure of Transformer into the graph.The model performs subgraph sampling on the dynamic network and uses Transformer to learn node and link embedding representations,enabling the model to better capture the node evolution patterns of the network.Compared with current state-ofthe-art algorithms,Dyn Trans has excellent performance in dynamic network node classification tasks and link prediction tasks.At the same time,its node embedding representation is more discriminative in high-dimensional space,showing outstanding performance on large-scale financial fraud network data. |
