Logo image
Combining Accuracy and Precision in Macsum Aggregation Learning
Acte de colloque

Combining Accuracy and Precision in Macsum Aggregation Learning

Isaac Kinane, Christophe Marsala, Agnès Rico et Olivier Strauss
Proceedings of the 12th International Conference on Soft Methods in Probability and Statistics
International Conference on Soft Methods in Probability and Statistics (Lecce, Italy, 15/09/2026–18/09/2026)

Résumé

Loss function Macsum aggregation learning
The macsum aggregation model has been recently introduced to generalise the concept of maxitive measure (possibility measure) to games. This model has several interesting features as the output is an interval instead of a precise value, and, although it is a game, it is based on only n parameters for n inputs, like a linear model.The parameters of a macsum model can be learned by linear regression from a training set of examples, however, in this case, the loss function to use must be choose carefully in order to optimise both the accuracy of the macsum output interval and its width. In this article, we propose a new loss function for this purpose, the Smooth Penalised Interval Loss, and present a series of experiments designed to demonstrate its effectiveness.

Fichiers et liens (1)

url
Find in HALAfficher

Indicateurs

1 Consultations de la notice

Détails

Logo image