Graph Convolutional Networks and Attention-Based Outlier Detection

Abstract

Outlier detection is a significant research direction in machine learning and has many applications in finance, network security, and other areas. Outlier detection of Euclidean datasets is a mainstream problem in outlier detection. Most detection methods often ignore the connection of its nodes. To collect the representation information of feature sets and node connections to improve the detection of outliers in Euclidean datasets Accuracy rate， we propose a novel Graph Convolutional and Attention-Based Outlier Detection (GCA).The GCA first converts the Euclidean structure data into directed graphs using locally sensitive hashing; then, by applying a Graph Convolutional Network, the data features and their connectivity graph are fed into the neural network; secondly, it fuses the extracted features and the features reconstructed by the attention mechanism; finally, calculating the outlier factors of the objects. Comparing eight state-of-art algorithms on ten real-world datasets shows that GCA achieves the highest Area Under ROC Curve (AUC) on datasets and also achieves equally good results in Accuracy (ACC) and False Alarm Rate (FAR). This study fills the gap of upgraded GCNs in detecting outliers to the best of our knowledge and provides a new way to convert Euclidean data to graphs.

Taxonomy of outlier detection methods for power system measurements

“…Learning-based methods learn from the data or their attributes. There are three different types of learning-based methods, viz., subspace learning, graph-based learning, and neural networkbased learning [78][79][80][81][82][83][84][85][86][87][88][89][90][91][92][93][94][95][96].…”

Health insurance fraud detection based on multi-channel heterogeneous graph structure learning

“…Nevertheless, there is room for improvement in both the model's generalization ability and the data transformation process. Qiu et al [18] introduced an outlier detection algorithm employing graph convolution and attention to tackle feature ambiguity and improve accuracy. While effective in numerous outlier detection tasks, its suitability for more complex scenarios may be limited.…”

“…In recent years, graph neural networks have garnered increasing attention in anomaly detection [16] , [17] , [18] . These methods represent the objects in the data as nodes, the relationships between them as edges, use graph neural networks to score anomalies appropriately, and identify data that may contain anomalous behavior.…”