Kernel self-representation residual analysis model for anomaly detection

Musheng Huang

doi:10.1016/j.asoc.2026.115517

Kernel self-representation residual analysis model for anomaly detection

2026-05-22

0

PRE

AI

DOI：10.1016/j.asoc.2026.115517

Original

Original request for help

Save

Abstract

En 中文

• This paper proposes KRadar model to address the limitation of existing linear residual models (Radar) in effectively capturing the nonlinear relationships among nodes. • The KRadar method projects the features of the attribute matrix and residual matrix into a high-dimensional Hilbert space through a nonlinear mapping, making normal and abnormal nodes linearly separable in the high-dimensional space. • We proposed a kernel-based self-representation model, which projects the attribute matrix X and the self-represented residual matrix R of nodes into a high-dimensional Hilbert space through kernel techniques. • Meanwhile, we introduced kernel techniques into the graph regularization term tr(RLRT), because the residual matrix R has been projected into a high-dimensional Hilbert space in the kernel-based self-representation model. • The experimental results compared to the other seven methods indicate that the average ROC-AUC value of KRadar is 10.2% higher than the suboptimal method Radar on two real datasets and six synthetic datasets.

Keywords:

KRadar

anomaly detection

kernel-based self-representation

residual analysis

Hilbert space

Kernel self-representation residual analysis model for anomaly detection

Abstract

Journal

Organization