Ensemble One-class Classification Based on BLS-Autoencoder

‍ ‍Anomaly detection is a classic research problem in the field of pattern recognition； however， it is difficult to apply conventional anomaly detection methods in the scenario of extreme class imbalance， where only normal samples are included in the training set. Thus， one-class classification （OC） has gained increasing interest owing to its ability to build decision boundaries for the target class to identify non-target samples. Although many state-of-the-art OC algorithms have been proposed， they still have some limitations： （1） They usually train the model in the original feature space， which can be easily affected by noise features； （2） Most of these methods design a single OC model， which makes it difficult to learn comprehensive decision boundaries from multiple perspectives of feature subspaces； （3） All training samples are treated equally， lacking a specific learning process for under-fitting samples. To address these problems， this paper proposes an ensemble one-class classification based on BLS-autoencoder （EOC-BLSAE）. Specifically， inspired by the advantages of good generalization performance of a broad learning system （BLS） and autoencoder， a one-class BLS-autoencoder （OC-BLSAE） is first designed. OC-BLSAE efficiently eliminates noise features by learning the nonlinear mapping relationship between the original feature space and the reconstructed feature space. The reconstruction error of training samples is utilized to establish the decision boundary for the target class， and the testing samples with large reconstruction errors are recognized as anomalies. Subsequently， to build OC-BLSAE models from multiple perspectives， a one-class boosting strategy was proposed， which iteratively produces OC-BLSAE models by feeding training subsets containing under-fitting samples. Theoretically， the sampling weights of previously under-fitting samples would increase if the overall reconstruction loss of multiple OC-BLSAE models was minimized， and the reliability of each OC-BLSAE model could be adaptively evaluated. Upon training multiple OC-BLSAE models， the predictions of these models are integrated by weighted voting to derive the final prediction. Parameter， comparison， significant test， and ablation experiments are performed on 16 different OC tasks to verify the effectiveness of the proposed methods. The experimental results demonstrate that the proposed methods could obtain better overall performance than state-of-the-art OC approaches.