As machine learning systems increasingly inform critical decisions in domains such as healthcare, finance, and risk management, understanding not only their predictive reliability but also the uncertainty surrounding their explanations becomes essential. This thesis investigates how different forms of uncertainty affect the interpretability and use of machine learning models in decision-making contexts. The research integrates analytical and behavioral perspectives across three studies. The first develops a methodological framework based on rank aggregation to reconcile divergent feature importance explanations arising from model multiplicity. The second proposes a decision framework to assist analysts in managing disagreement among models, translating analytical scenarios into practical strategies for model selection and interpretation. The third examines, through two behavioral experiments, how predictive uncertainty influences human reliance on algorithmic and hybrid advice. Overall, the thesis contributes to bridging the gap between explainable artificial intelligence and decision-making under uncertainty.
Con il crescente impiego dei sistemi di machine learning nelle decisioni critiche in ambiti come la sanità, la finanza e la gestione del rischio, comprendere non solo la loro affidabilità predittiva, ma anche l’incertezza che circonda le loro spiegazioni diventa essenziale. Questa tesi indaga come diverse forme di incertezza influenzino l’interpretabilità e l’utilizzo dei modelli di machine learning nei contesti decisionali. La ricerca integra prospettive analitiche e comportamentali attraverso tre studi. Il primo sviluppa un framework metodologico basato sull’aggregazione di ranking per conciliare spiegazioni divergenti dell’importanza delle variabili dovute dalla presenza di una molteplicità di modelli accurati. Il secondo studio propone un framework decisionale per supportare gli analisti nella gestione delle divergenze tra modelli, traducendo scenari analitici in strategie pratiche per la selezione e l’interpretazione dei modelli. Il terzo studio esamina, attraverso due esperimenti comportamentali, come l’incertezza predittiva influenzi la fiducia umana nei suggerimenti forniti da algoritmi e sistemi ibridi. Nel complesso, la tesi contribuisce a colmare il divario tra intelligenza artificiale spiegabile e decision-making in condizioni di incertezza.
Artificial uncertainty
SESSA, Claudia
2025
Abstract
As machine learning systems increasingly inform critical decisions in domains such as healthcare, finance, and risk management, understanding not only their predictive reliability but also the uncertainty surrounding their explanations becomes essential. This thesis investigates how different forms of uncertainty affect the interpretability and use of machine learning models in decision-making contexts. The research integrates analytical and behavioral perspectives across three studies. The first develops a methodological framework based on rank aggregation to reconcile divergent feature importance explanations arising from model multiplicity. The second proposes a decision framework to assist analysts in managing disagreement among models, translating analytical scenarios into practical strategies for model selection and interpretation. The third examines, through two behavioral experiments, how predictive uncertainty influences human reliance on algorithmic and hybrid advice. Overall, the thesis contributes to bridging the gap between explainable artificial intelligence and decision-making under uncertainty.| File | Dimensione | Formato | |
|---|---|---|---|
|
SessaClaudia.pdf
non disponibili
Licenza:
Tutti i diritti riservati
Dimensione
4.33 MB
Formato
Adobe PDF
|
4.33 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/385850
URN:NBN:IT:LIUC-385850