Multisphere Support Vector Data Description: A Mathematical Optimization Approach

V. Blanco Izquierdo, I. Espejo Miranda, R. Páez Jiménez, A. M. Rodriguez Chía

In this work, we present a mathematical optimization-based approach for developing an outlier detection tool for multimodal datasets. We begin by deriving a primal problem to compute Euclidean hyperspheres for identifying outlier observations. We then develop a dual formulation that enables the use of the kernel trick, allowing for the creation of more complex boundaries for regular observations. We report the results of extensive computational experiments that validate our proposal compared to previous heuristic approaches in the literature.

Palabras clave: Multisphere SVDD, multimodal outlier detection, kernel trick

Localización (GELOCA1)
11 de junio de 2025  15:30
MR 2

P. Segura Martínez, A. Marín

F. Temprano Garcia, L. Mallozzi, J. Puerto

S. Anglada, M. C. Galé Pola, J. J. Salazar González

V. Blanco