logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-08312023-150258


Tipo di tesi
Tesi di laurea magistrale
Autore
NOCCHI, TOMMASO
URN
etd-08312023-150258
Titolo
Design and Experimental Evaluation of a Novel Model-Agnostic Feature Importance Measure for Quality Measures in Industrial Production Processes
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
ARTIFICIAL INTELLIGENCE AND DATA ENGINEERING
Relatori
relatore Ing. Alfeo, Antonio Luca
relatore Prof. Cimino, Mario Giovanni Cosimo Antonio
relatore Ing. Gagliardi, Guido
Parole chiave
  • AI
  • counterfactual
  • counterfactual explanation
  • Explainable Artificial Intelligence
  • feature importance
  • XAI
Data inizio appello
22/09/2023
Consultabilità
Tesi non consultabile
Riassunto
Industry 4.0 signifies the current industrial revolution, leveraging modern technologies to establish intelligent environments for enhanced production processes. The product quality is closely linked to the manufacturing decision system configuration. Predicting product quality could optimize parameters early in design, boosting competitiveness. While machine learning aids in prediction, its opacity hinders comprehension in critical industrial contexts. Explainable AI (XAI) emerges, with counterfactual-based explanations being effective in revealing input-output relationships. DiCE is a XAI framework generating counterfactuals to elucidate model predictions, offering diverse explanations by altering inputs. Proposed XAI doesn't directly use counterfactuals but derives feature importance using counterfactual samples for a specific instance. This leads to a global explanation termed BoCSoR 2.0, by quantifying the frequency with which a change in each feature in isolation lead to a significant change in the regressor model’s output. Comparisons between DiCE and BoCSoR 2.0 performances are organized. Experimental results on synthetic and industrial data support BoCSoR 2.0's reliability, agreement with expert knowledge, and efficiency. BoCSoR 2.0 proves promising for industry due to its effectiveness, efficiency, and robustness to feature correlation, making it a valuable XAI tool.
File