An Unsupervised Learning-Based Multivariate Anomaly Detection Method for Dynamic Attention Graphs

The widespread adoption of information technology has led to the proliferation of data applications, underscoring the importance of identifying potential anomalies within datasets. However, developing anomaly detection models for intricate data presents a formidable challenge. Issues such as the scarcity and inferior quality of anomaly data hinder the effectiveness of supervised models, necessitating the exploration of unsupervised detection methods. This manuscript proposes an unsupervised learning-based multivariate anomaly detection method for dynamic attention graphs (ADDAG), which combines the characteristics of graph convolutional neural networks that fuse the relationships between data while updating their own nodes to reconstruct the data and extract the anomalous data. Compared with the traditional static attention graph, the use of dynamic attention graph enhances the ability to express the features of complex associated time-series variables. In the experimental part, the performance of the proposed method in this paper outperforms conventional anomaly detection algorithms when tested on multiple datasets.