A robust multi-view support vector machine with the RoBoSS loss function

Abstract

Multi-view learning capitalizes the multiple representations of data and integrates their heterogeneous information to improve the learning performance. Several support vector machine (SVM)-based multi-view models have been proposed and shown excellent performance. However, they mainly rely on consensus-based strategies and neglect the complementary information. Moreover, these approaches lack robustness against noise, errors, and view-inconsistent patterns, common in multi-view datasets. To overcome these limitations, we propose a robust multi-view SVM framework that utilizes the RoBoSS loss function called RoBoSS-MvSVM. The proposed method explicitly integrates consensus and complementarity information across views, enriching the multi-view data representation and ensuring resilient learning. The RoBoSS loss function exhibits robustness, boundedness, sparsity, and smoothness, which makes it effective in handling noisy and inconsistent samples, while its classification-calibrated theoretical property ensures reliable generalization performance. We use the Nesterov accelerated gradient algorithm to solve the optimization problem of the proposed RoBoSS-MvSVM. Furthermore, the generalization capacity of the proposed RoBoSS-MvSVM technique is theoretically established through Rademacher complexity analysis. To further validate its robustness and effectiveness in multi-view learning, comprehensive experiments are conducted on 3 synthetic, 39 benchmark UCI and KEEL, and 45 Animal with Attribute datasets. The experimental results consistently demonstrate that RoBoSS-MvSVM outperforms existing baseline methods. Additionally, hyperparameter sensitivity analysis and statistical evaluation confirm the stability and significance of the proposed model’s generalization performance. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.