EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2965. Enhancing interpretability in Twin Support Vector Machines via variable selection

Invited abstract in session WB-27: Unraveling the Black Box: Advances in Model Explainability, stream Mathematical Optimization for XAI.

Wednesday, 10:30-12:00
Room: 047 (building: 208)

Authors (first author is the speaker)

1. Sandra Benítez-Peña
Statistics Department, Universidad Carlos III de Madrid

Abstract

Support Vector Machines (SVM) constitute a highly effective tool for solving binary clasification problems. On the other hand, Twin Support Vector Machines (TWSVM) are based on the idea of constructing SVMs in pairs through generalized eigenvalues (GEPSVM). These determine two hyperplanes not necessarily parallel by solving two problems similar to SVM but of smaller size. Consequently, the computational cost during the training phase is significantly reduced compared to traditional SVM. This requires solving two smaller quadratic programming problems (QPPs) that the one solved by the standard SVM.
Moreover, many real-life problems, such as those related to fraud prediction and credit scoring, involve costs of misclassification that can be different for both classes. Although providing accurate values for such misclassification costs can be challenging for the user, identifying acceptable misclassification rates is relatively straightforward. In this work, we propose a new TWSVM in which misclassification costs are considered by incorporating performance constraints into the problem formulation. This model is further enriched by performing variable selection (FS), a fundamental task that makes the method more interpretable and efficient. Numerical results demonstrate the effectiveness of the proposed method.

Keywords

Status: accepted

Back to the list of papers