Orador Convidado: Pedro Silva (Universidade Católica do Porto)

Resumo: We propose a Support Vector Machine (SVM) methodology for finding accurate probabilities of memberships in supervised classification problems. Classical SVMs do not complement their prediction with reliable confidence measures for each assignment. While classical statistical methodologies can address this problem, they tend to rely on restrictive assumptions that are often difficult to guarantee in practice. Following a fully distribution-free non-parametric approach, we show how, first two-class classification problems, sequences of weighted SVM predictions can be used to consistently estimate probabilities of membership. This methodology is then extended to the general multi-class classification problem. Numerical experiments reveal the good scaling properties of the proposal, and the relative advantages of its probability estimates over alternative approaches.

Esta comunicação será, em parte, baseada no seguinte artigo, recente publicado na revista “Computers and Operations Research”

Duarte Silva A.P. (2025), “Probabilistic Vector Machines”, Computers and Operations Research, Vol 183, https://doi.org/10.1016/j.cor.2025.107203