ScholarWorks@숙명여자대학교

초록

Recently, anomaly detection methods for graph nodes have been actively researched in various fields. Most of the previous studies designed models using graph neural networks (GNNs) to detect outliers. Anomaly detection for graph nodes aims to identify rare nodes that differ from the majority of nodes by learning the characteristics of the entire graph or each node. This paper specifically focuses on the characteristics of social network graphs. Most of the existing outlier detection studies compare the usage patterns of regular users with those of anomalous users, or outliers. Graph-based methods propose more effective detection methods by utilizing the usage patterns of anomalous users. However, there is a limitation in that the characteristic information of each node is extracted using only structured data. This paper proposes a technique for detecting outlier nodes in a graph considering their text attributes. In this paper, an extended feature matrix is created that can more accurately predict outlier detection when learning the pattern of each node. Specifically, Word2vec is used to generate embedding vectors, which are then added to the extended feature matrix to evaluate the performance improvement of GNN. Word2vec is a natural language processing model mainly used for generating word embeddings. The transformed embedding vectors are added to the feature matrix to detect values that the existing model cannot restore well as anomalies. In this paper, virtual data consisting of mixed normal and anomalous data were used. As a result of applying the proposed method, it resulted in a high F1-score of 98% and an AUC of 99.3%.

키워드