Title of the talk: Enhancing robustness and sparsity: Least squares one-class support vector machine

Date & Time: Nov 20,2025 at 12 noon  (ONLINE)

Abstract: Identifying unusual or abnormal data, often called one-class classification (OCC) is a fundamental problem across science and engineering. From detecting faults in systems to identifying rare events in high-dimensional datasets, OCC plays a central role in many real-world applications. A well-known model for OCC is the least squares one-class support vector machine (LS-OCSVM), valued for its simplicity and elegant mathematical formulation. However, the classical model faces two major challenges: First, it is sensitive to outliers and noise. Second, it produces non-sparse solutions, which limits its scalability to large problems. To address these limitation, we proposed two models. First, robust LS-1SVM (RLS-1SVM),which improves stability by controlling both the mean and variance of the error, making the model resistant to noisy data. Second, sparse robust LS-1SVM (SRLS-1SVM), which introduces sparsity through low-rank approximations, creating the first sparse extension of LS-OCSVM suitable for large-scale. The work includes intuitive explanations, highlighting the mathematical ideas behind robustness, sparsity, and generalization. Further, practical advantages, and empirical results are presented that demonstrate the improved performance of our models on benchmark datasets.

About the speaker: Dr. Anuradha Kumari obtained her Master’s degree in Mathematics and Computing from the Indian Institute of Technology Guwahati (IITG) in 2020. Subsequently, she pursued her Ph.D. in Mathematics at the Indian Institute of Technology Indore (IITI), which she completed in May 2025. Currently, Dr. Kumari is a Postdoctoral Researcher at KU Leuven, Belgium, working under the joint supervision of Prof. Johan Suykens and Dr. Panagiotis Patrinos on the project titled “Rethinking Transformers Through Duality”. Her broader research interests include kernel methods, robust machine learning, one-class classification, ensemble deep randomized neural networks and attention mechanisms in transformers.