Proposal Description

This IPS focuses on most relevant methodological and applied issues of data science and artificial intelligence with particular attention to interpretability of results and potential bias. The viewpoint of computer scientists and engineers is given to strengthen the collaboration between statisticians and other communities in the field of data science. This IPS is balanced from geographical and gender point of view and is proposed by the ISI Special Interest Group on Data Science.
Francesco Sovrano Title: Explainability for Software Engineering in the Age of LLMs
This talk examines how statistical biases in large language models (LLMs) used for software engineering undermine reliability, and how explainability can help uncover and mitigate them. Focusing on two representative tasks (code clone detection and vulnerability localization) it first shows that current LLM-based systems often overfit standard benchmarks, leading to sharp performance drops on out-of-distribution code and revealing hidden training and evaluation biases. In vulnerability localization, the models learn harmful shortcuts (e.g. “bugs are usually early in short files”), causing them to systematically miss flaws in later parts of long files and to trigger excessive false positives when naive chunking is applied. To study such behaviours at scale, the talk proposes an abstraction-driven explainable AI (XAI) framework: instead of perturbing raw code, it analyses model behaviour over higher-level properties and uses injected ground-truth biases to benchmark XAI methods. Building on this, it introduces RuleSHAP, which combines rule learning with SHAP-based sparsification to better recover nonlinear, conjunctive biases, and outlines open challenges for bias detection, mitigation, and communication in safety-critical, code-facing LLMs.
Fatima Rabia Yapicioglu and Alberto Rigenti Title: XAI Applications in Business Intelligence and Data Science for the Super Sport Vehicle Industry
This talk explores how Explainable AI supports data-driven decision making within the super sport vehicle industry. We will present practical applications of Conformal Prediction, customer-journey analytics, vehicle-ownership tracing, and customer-scoring models, and discuss how XAI methods help ensure transparency, reliability, and compliance with emerging legal frameworks governing customer–vehicle data relationships.
Helena Löfström-Cavallin Title: Ensured Explanations: Uncertainty-Aware Reasoning in Explainable Data Science
This talk introduces a framework that expands the traditional one-dimensional view of explaining predicted probabilities into a two-dimensional space of probability and uncertainty, revealing how explanations can meaningfully move models across both axes rather than along a single confidence scale. The framework categorizes explanation outcomes: ensured, semi-potential, counter-potential, and super-potential, based on how they shift a prediction’s position within this expanded space. The perspective makes explicit the consequences of explanation-driven changes: an explanation may increase or decrease predicted probability, reduce or amplify uncertainty, or do both in different combinations, enabling a more structured and operational understanding of explanation effects. To demonstrate this, the talk presents an extension of Calibrated Explanations with a tunable scoring metric that allows practitioners to prioritize uncertainty reduction, confidence adjustment, or a deliberate trade-off. Overall, the ensured explanation framework reframes explainability as an active tool for uncertainty-aware data analysis, highlighting how explanations can guide models toward more reliable regions of the probability–uncertainty space and support trust-focused decision-making in complex workflows.