logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-11022022-103624


Tipo di tesi
Tesi di laurea magistrale
Autore
MARTORANA, MICHELANGELO
Indirizzo email
m.martorana1@studenti.unipi.it, mhlmrt@gmail.com
URN
etd-11022022-103624
Titolo
A novel feature importance measure to explain the quality level prediction in Smart Manufacturing
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Relatori
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Prof. Alfeo, Antonio Luca
relatore Ing. Gagliardi, Guido
Parole chiave
  • Bocsor
  • Contrastive learning
  • Counterfatuals
  • Feature importance
  • industrial data
  • predictive maintenance
  • smart manifacturing
Data inizio appello
18/11/2022
Consultabilità
Non consultabile
Data di rilascio
18/11/2025
Riassunto
The new feature importance approach, called Bocsor, is based on counterfactuals searching. For the approach, it is necessary to explain the decision boundary between two classes by highlighting the most relevant features for moving from one class to another. To do so is possible to focus on the samples that are closest to the decision boundary. Starting with the counterfactual is possible to replace the feature values with those of the original sample. If this substitution corresponds to passing through the decision boundary, that feature is considered relevant, since it alone allows one to cross the decision boundary. By repeating this procedure for each sample close to the boundary, it is possible to count how often each feature alone allows the decision boundary to pass through when the samples are in proximity to it. Since it is evaluated, once we are close to the decision boundary, the change of each feature by itself results in the crossing of the model’s decision boundary, we call this measure the Boundary Crossing Solo Ratio (BoCSoR). The thesis is based on the design and experimental of this new feature importance approach compared to the actual state of the art of feature importance approaches.
File