The thesis starts by arguing that current hybrid decision-making systems rely on a mutually exclusive division of responsibilities between human and machine decision makers. We argue that this phenomenon is suboptimal as most of the typically used frameworks leave one of the partners completely blind while a decision is being made, resulting in the system being less trustworthy and accountable. This thesis proposes a solution to this problem by introducing Learning to Guide (LtG), where both the user and the machine contribute to the decision-making task. Furthermore, we advocate that for high-stakes decision making tasks, the right to make the downstream decision should be reserved for the human. Therefore, the objective is to develop a framework that instead of being a simple decision-maker, provides guidance to the user, essentially reserving the scope of decision-making to the end-user. We first developed Slog, a novel framework that operates on top of any model, assigned to certain task. In this task, given an image, Slog provides textual guidance to the end user in order to help the latter in decision making tasks. The utility of Slog was evaluated with a synthetic dataset and a challenging medical dataset, along with experiments involving human users. During this exercise we noticed that fine-tuning a model can be computationally expensive. Hence the need for a fine-tuning free inferential framework was evident. As a remedy, we introduced MedGellan, a clinically specialized framework that assists the physician in clinical diagnostics. Nevertheless, both Slog and MedGellan share a common limitation: the guidance generated compounds information across all possible outcomes, which can be difficult for the expert to digest — particularly in multi-label settings where the user expects a focused and targeted set of predictions rather than a comprehensive overview. This observation motivated our third contribution, ConfGuide, a novel learning-to-guide approach that generates more succinct and targeted guidance by employing conformal risk control to select a focused set of outcomes. Crucially, ConfGuide provides formal statistical guarantees on the false negative rate — ensuring that likely present pathologies are not overlooked — a property that is especially desirable in high-stakes clinical settings where the cost of a missed diagnosis far outweighs the cost of a false alarm. Under a simulated experimental setup, ConfGuide demonstrated superior performance over its competitors, further reinforcing the central claim of this thesis: that equipping the human expert with targeted, principled guidance — rather than autonomous machine decisions — is a more effective and trustworthy paradigm for high-stakes decision making.

Building Machines that Learn to Guide Human Decision Makers

BANERJEE, DEBODEEP
2026

Abstract

The thesis starts by arguing that current hybrid decision-making systems rely on a mutually exclusive division of responsibilities between human and machine decision makers. We argue that this phenomenon is suboptimal as most of the typically used frameworks leave one of the partners completely blind while a decision is being made, resulting in the system being less trustworthy and accountable. This thesis proposes a solution to this problem by introducing Learning to Guide (LtG), where both the user and the machine contribute to the decision-making task. Furthermore, we advocate that for high-stakes decision making tasks, the right to make the downstream decision should be reserved for the human. Therefore, the objective is to develop a framework that instead of being a simple decision-maker, provides guidance to the user, essentially reserving the scope of decision-making to the end-user. We first developed Slog, a novel framework that operates on top of any model, assigned to certain task. In this task, given an image, Slog provides textual guidance to the end user in order to help the latter in decision making tasks. The utility of Slog was evaluated with a synthetic dataset and a challenging medical dataset, along with experiments involving human users. During this exercise we noticed that fine-tuning a model can be computationally expensive. Hence the need for a fine-tuning free inferential framework was evident. As a remedy, we introduced MedGellan, a clinically specialized framework that assists the physician in clinical diagnostics. Nevertheless, both Slog and MedGellan share a common limitation: the guidance generated compounds information across all possible outcomes, which can be difficult for the expert to digest — particularly in multi-label settings where the user expects a focused and targeted set of predictions rather than a comprehensive overview. This observation motivated our third contribution, ConfGuide, a novel learning-to-guide approach that generates more succinct and targeted guidance by employing conformal risk control to select a focused set of outcomes. Crucially, ConfGuide provides formal statistical guarantees on the false negative rate — ensuring that likely present pathologies are not overlooked — a property that is especially desirable in high-stakes clinical settings where the cost of a missed diagnosis far outweighs the cost of a false alarm. Under a simulated experimental setup, ConfGuide demonstrated superior performance over its competitors, further reinforcing the central claim of this thesis: that equipping the human expert with targeted, principled guidance — rather than autonomous machine decisions — is a more effective and trustworthy paradigm for high-stakes decision making.
13-nov-2026
Inglese
clinical diagnosis
conformal prediction
conformal risk control
hybrid decision making
large language model
medical decision-making
vision language model
Passerini, Andrea
Teso, Stefano
File in questo prodotto:
File Dimensione Formato  
thesis_final_Banerjee_pdfa_compressed.pdf

accesso aperto

Licenza: Creative Commons
Dimensione 6.18 MB
Formato Adobe PDF
6.18 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/385762
Il codice NBN di questa tesi è URN:NBN:IT:UNIPI-385762