چكيده به لاتين
In recent years, graph data has received a lot of attention; Because they are also used to represent other types of data, including social network, banking, security, financial, medical, and textual data. Therefore, the detection of anomalies in these data has received increasing attention due to their unfortunate consequences and has shown its power in preventing destructive events such as financial fraud, network intrusion, and social spam. In general, anomalies are patterns in data that do not conform to the defined concept of normal behavior. Such anomalies in graph data may be seen in several ways: 1- Anomaly in the node, 2- Anomaly in the edge, 3- Anomaly in the subgraph, 4- Anomaly in the graph. A nodal abnormality may be due to abnormal structure or features or both. On the other hand, edge-related anomalies, unlike node anomaly detection, which targets individual nodes, the purpose of edge anomaly detection is to identify abnormal links. These links are often unexpected or unusual relationships between real objects, such as unusual interactions between fraudsters and benign users, or suspicious interactions between attacker nodes and benign user machines in computer networks. The main challenge in this field is to identify these abnormalities and classify them. In recent years, many computational methods have been developed to predict anomalies in graphs. These methods can detect anomalies in the graph. These calculation methods are generally divided into two categories based on statistical analysis and based on machine learning.
In this research, through the application of anomaly detection methods at the node level, we present "an anomaly identification framework at the node level" with the approach of deep and ensemble learning algorithms. In this research, it was tried to apply the proposed framework on labeled hetrogeneous graphs and compare its results with other algorithms for identifying anomalous nodes quantitatively and qualitatively.
A quantitative evaluation of the "GRAPH-Guard" framework on heterogeneous labeled graphs showed that using the proposed framework results in increased accuracy, based on the AUC metric (+4) compared to previous works' average and (+1) F1 compared to the best value in previous works
