Machine learning (ML) holds great promise for transforming healthcare, by enhancing diagnostic accuracy, enabling earlier interventions, and refining treatment strategies. Yet, the opacity of many ML models remains a major barrier to clinical adoption. Explainable Artificial Intelligence (XAI) is often proposed as a way to address this challenge, but we still know little about how different stakeholders perceive and engage with XAI in practice. This study explored how three XAI techniques, Local SHAP, Global SHAP, and Attention Mechanism, shape trust, usability, and decision-making across diverse user groups. We conducted semi-structured interviews with 20 participants, including clinical experts, data scientists, and members of the public, using real-world ML prediction tasks in two healthcare domains: sepsis-related mortality and psychology. Participants reviewed visualizations of model explanations and shared their perspectives in guided discussions. Through thematic analysis, five key themes emerged. Participants stressed that explainability must be paired with perceived reliability, contextual relevance, and human oversight to build trust. Local SHAP was generally seen as the most intuitive method, while Global SHAP and Attention Mechanism were often perceived as too abstract or technical. Across all groups, AI was viewed as a decision-support tool, not a replacement for clinical judgment. Successful integration of XAI was seen as dependent on adaptable explanations, integration into existing workflows, appropriate training, and interdisciplinary collaboration. Our findings highlight the need for human-centered design and institutional readiness in the development of trustworthy, usable XAI systems for healthcare.
Evaluating machine learning (ML) model bias is key to building trustworthy and robust ML systems. Counterfactual Fairness (CF) audits allow the measurement of bias of ML models with a causal framework, yet their conclusions rely on a single causal graph that is rarely known with certainty in real-world scenarios. We propose CF with Graph Uncertainty (CF-GU), a bias evaluation procedure that incorporates the uncertainty of specifying a causal graph into CF. CF-GU (i) bootstraps a Causal Discovery algorithm under domain knowledge constraints to produce a bag of plausible Directed Acyclic Graphs (DAGs), (ii) quantifies graph uncertainty with the normalized Shannon entropy, and (iii) provides confidence bounds on CF metrics. Experiments on synthetic data show how contrasting domain knowledge assumptions support or refute audits of CF, while experiments on real-world data (COMPAS and Adult datasets) pinpoint well-known biases with high confidence, even when supplied with minimal domain knowledge constraints.
Large Language Models (LLMs) are known to overuse certain terms like "delve" and "intricate." The exact reasons for these lexical choices, however, have been unclear. Using Meta's Llama model, this study investigates the contribution of Learning from Human Feedback (LHF), under which we subsume Reinforcement Learning from Human Feedback and Direct Preference Optimization. We present a straightforward procedure for detecting the lexical preferences of LLMs that are potentially LHF-induced. Next, we more conclusively link LHF to lexical overuse by experimentally emulating the LHF procedure and demonstrating that participants systematically prefer text variants that include certain words. This lexical overuse can be seen as a sort of misalignment, though our study highlights the potential divergence between the lexical expectations of different populations -- namely LHF workers versus LLM users. Our work contributes to the growing body of research on explainable artificial intelligence and emphasizes the importance of both data and procedural transparency in alignment research.
Missing data represents one of the most ubiquitous data quality issues, and also one of the most impactful on machine learning (ML) pipelines. Indeed, not only most commonly applied ML methods cannot directly employ incomplete data, but also the techniques employed to manage this issue can impact on the performance and evaluation of ML models. Among such techniques to manage missing data, imputation, that is filling in the missing values using information from the observed data, remains among the most popular and effective in practice. Yet, from a theoretical point of view, it is still not clear under which conditions it is possible to learn effectively after imputation. In this article we address this gap by studying learnability under imputation in the framework of statistical learning theory. After giving a general definition of learnability under imputation, we show three main contributions: 1) we introduce a novel stability condition, called noise risk stability, which we prove to be both sufficient and, under weak assumptions, necessary for learnability under imputation; 2) we show that a large class of ML models (including linear and kernel methods) satisfies noise risk stability; 3) we characterize the learning-theoretic properties of two common imputation methods (constant and regression imputation). Our results set the stage for a rigorous study of imputation and missing data management in the framework of statistical learning theory, by also describing relevant open questions.
The goal of learning to hash (L2H) is to derive data-dependent hash functions from a given data distribution in order to map data from the input space to a binary coding space. Despite the success of L2H, two observations have cast doubt on the source of the power of L2H, i.e., learning. First, a recent study shows that even using a version of locality sensitive hashing functions without learning achieves binary representations that have comparable accuracy as those of L2H, but with less time cost. Second, existing L2H methods are constrained to three types of hash functions: thresholding, hyperspheres, and hyperplanes only. In this paper, we unveil the potential of Voronoi diagrams in hashing. Voronoi diagram is a suitable candidate because of its three properties. This discovery has led us to propose a simple and efficient no-learning binary hashing method, called Voronoi Diagram Encoded Hashing (VDeH), which constructs a set of hash functions through a data-dependent similarity measure and produces independent binary bits through encoded hashing. We demonstrate through experiments on several benchmark datasets that VDeH achieves superior performance and lower computational cost compared to existing state-of-the-art methods under the same bit length. Our code is available at: https://github.com/Phantom-Det /VDeH/tree/main.
