Verfahren zur Gradientbasierten Bayesischen Informationskriteriumskompression von Entscheidungsbäumen

Details

Veröffentlichungs-Nr.

EP4749535

Anmeldetag

21. November 2024

Veröffentlichung

27. Mai 2026

Rechtsraum

EP

IPC

G06N5/01, G06N7/01, G06N20/20

Abstract

A computer-implemented method of compressing a decision tree by pruning its structure, wherein the decision tree is configured to determine a classification of its input, the method comprising the steps of: Receiving trained decision tree comprising a plurality of nodes, wherein each inner node is associated with a branching function. Outputting a score for each class label based on the traversed paths and the outputs of the reached leaf nodes and inner leaf nodes. Determining based on the outputs and the parameters of the Decision Tree a Bayesian Information Criterion (BIC). Optimizing at least the parameter of the probabilistic branching functions and the first parameters by minimizing a loss function based on the Bayesian Information Criterion (BIC), wherein the optimization is performed using gradient descent.

Anmelder

Firma

Robert Bosch GmbH

Land

🇩🇪 Deutschland