EDA-OCBLS: An error-distribution aware one-class broad learning system for anomaly detection

Abstract

One-class classification (OCC) has emerged as a fundamental paradigm for detecting anomalous patterns and distributional shifts in real-world data. However, existing models often exhibit instability when exposed to noisy or contaminated environments. The recently proposed one-class broad learning system (OCBLS) offers a scalable flat-architecture solution with efficient feature mapping and representational capacity, but its sensitivity to outliers and distributional irregularities severely undermines its generalization capability. To overcome these limitations, we propose the error-distribution aware one-class broad learning system (EDA-OCBLS), a novel model that introduces a variance-regularized loss function for distributional stability. Specifically, EDA-OCBLS simultaneously optimizes the first-order (mean) and second-order (variance) statistics of predictive errors, rather than relying solely on conventional l2-based reconstruction loss. This dual-objective error modeling mitigates the undue influence of outlier instances, enforces statistical stability in the latent representation space, and preserves structural consistency across feature dimensions. The improved robustness of the proposed model is formally established through rigorous theoretical analysis. Furthermore, extensive experiments on UCI and KDD benchmark datasets demonstrate the effectiveness of the proposed method. EDA-OCBLS achieves average accuracy improvements of approximately 4%-5% on UCI datasets and 5%-6% on KDD datasets, reaching an average accuracy of 99.69%. The proposed model EDA-OCBLS also shows strong robustness under noise while maintaining competitive computational efficiency. These improvements are further supported by statistical analyses, confirming the significance and consistency of the results. Collectively, these findings establish the proposed EDA-OCBLS model as a robust and effective solution for anomaly detection. Our code is publicly available https://github.com/mtanveer1/EDA-OCBLS © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.