Normal‐Mode‐Aware Augmented Kernel One‐Class SVM for Heterogeneous Normality Modeling

Despite the rapid advances in anomaly detection, many existing methods assume that normal data are homogeneous and can be modeled by a single one‐class boundary. However, in many practical scenarios, normal data are heterogeneous and consist of multiple latent modes, making conventional latent‐space kernels insufficient to characterize normality. To address this issue, we propose a normal‐mode‐aware augmented‐kernel one‐class support vector machine (One‐Class SVM, OCSVM) for heterogeneous normality modeling. A shared encoder maps inputs into a latent representation, while a gate head estimates the probabilistic affiliation of each sample using normal‐class labels to supervise heterogeneous normality modeling. The learned latent representation and gate probability are then combined through an augmented kernel, which measures both latent feature similarity and normal‐mode affiliation similarity. A unified one‐class SVM is trained with the resulting normal‐mode‐aware kernel. We evaluate the proposed method on several UCI tabular datasets and the MNIST benchmark. Experimental results show that the proposed method effectively captures heterogeneous normal patterns and improves anomaly detection performance across both tabular and image‐based settings. © 2026 The Author(s). IEEJ Transactions on Electrical and Electronic Engineering published by Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.