Abnormal User Detection via Multiview Graph Clustering in the Mobile e-Commerce Network
Table 3
Reference models used in the experiments.
Models
Model description
IF
Isolation Forest (IF) [44] is a tree-based anomaly detection model, which assumes that the abnormal objects can be isolated from others by fewer random feature segmentations compared with normal objects.
KNN
The K-nearest neighbor (KNN) model [45] recognizes the outliers as abnormal objects by comparing distances between objects.
LOF
The local outlier factor (LOF) algorithm [46] is based on an assumption that the local density of a normal object should be close to its neighbor’s density, while the local density of an abnormal object will be remarkably different from its neighbor’s density.
CAE+IF/CAE+KNN/CAE+LOF
These models are all composed of two components: a convolutional autoencoder is used to obtain low-dimensional embeddings, and an abnormal detector (IF/KNN/LOF) is used to discover abnormal objects based on the embeddings.
DeepFD
The model encodes the user-item bipartite graph into low-dimensional user representations with behavioral features using an autoencoder and employs DBSCAN to detect fraud block based on the representations [22].
FraudNE
The model captures the high-nonlinear characteristics from the user-item bipartite graph by an autoencoder and recognizes multiple fraudulent groups by predicting the cluster assignments based on the user representations [23].