Position Paper - Bridging the Gap Between Machine Learning and Sensitivity Analysis - Unified SA-based view of ML explanations

Aug 5, 2026

The paper presents a unified sensitivity analysis (SA) perspective on machine learning (ML) interpretations and model-building. It argues that explanations in ML and the broader ML process can be formalized as system-level SA, linking interpretable ML approaches and hyperparameter optimization to established SA methodology. By comparing interpretation methods with SA constructs and outlining how SA techniques can be transferred to ML, the work highlights benefits of a common framework and emphasizes fully crediting related research across fields.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Introduction

Why This Position Paper

Our Position

Position Paper: Bridging the Gap Between Machine Learning and Sensitivity Analysis Christian A. Scholbeck 1 2 3 Julia Moosbauer 1 2 Giuseppe Casalicchio 1 2 Hoshin Gupta 3 Bernd Bischl 1 2 Christian Heumann 1 arXiv :2312 . 13234v2 [ cs .LG] 10 Sep 2024 Abstract We argue that interpretations of machine learning (ML) models or the model-building process can be seen as a form of sensitivity analysis (SA), a general methodology used to explain complex systems in many fields such as environmental modeling, engineering, or economics. We address both researchers and practitioners, calling attention to the benefits of a unified SA-based view of explanations in ML and the necessity to fully credit related work. We bridge the gap between both fields by formally describing how (a) the ML process is a system suitable for SA,(b) how existing ML interpretation methods relate to this perspective, and (c) how other SA techniques could be applied to ML. 1. Introduction Machine learning (ML) is concerned with learning models from data with applications as diverse as text (Zhang et al., 2015) and speech processing (Bhangale & Mohanaprasad, 2021), robotics (Pierson & Gashler, 2017), medicine (Rajkomar et al., 2019), climate research (Rolnick et al., 2022), or finance (Huang et al., 2020) . Due to the increasing availability of data and computational resources, demand for ML has risen sharply in recent years, permeating all aspects of life. While the first publications in predictive modeling date back as far as the 1800s with Gauß and Legendre (Molnar et al., 2020 ; Stigler, 1981), the popularity of ML has surged in the twenty-first century, as it represents the current technological backbone for artificial intelligence. Increasing focus is put on interpretable models or the interpretation of blackbox models with model-agnostic techniques (Molnar, 2022 ; Rudin et al., 2022), often referred to as interpretable ML 1Department of Statistics, Ludwig-Maximilians-Universittin Munich, Munich, Germany 2Munich Center for Machine Learning (MCML), Munich, Germany 3Department of Hydrology and Atmospheric Sciences, The University of Arizona, Tucson AZ, USA. Correspondence to: Christian A. Scholbeck . (IML) or explainable artificial intelligence. Note that we utilize the term black box, although the internal workings of a model may be accessible but too complex for the human mind to comprehend. Furthermore, interpretations of the hyperparameter optimization (HPO) process have garnered attention in recent years (Hutter et al., 2014) . In the context of this paper, we will refer to IML as any effort to gain an understanding of ML, including HPO. In a basic sense, sensitivity analysis (SA) (Saltelli et al., 2008 ; Razavi et al., 2021 ; Iooss & Lema, 2015) is the study of how model output is influenced by model inputs. It is used as an assistance in many fields to explain inputoutput relationships of complex systems. Applications include environmental modeling (Song et al., 2015 ; Wagener & Pianosi, 2019 ; Shin et al., 2013 ; Haghnegahdar & Razavi, 2017 ; Gao et al., 2023 ; Mai et al., 2022 ; Nossent et al., 2011), biology (Sumner et al., 2012), engineering (Guo et al., 2016 ; Ballester-Ripoll et al., 2019 ; Becker et al., 2011), nuclear safety (Saltelli & Tarantola, 2002), energy management (Tian, 2013), economics (Harenberg et al., 2019 ; Ratto, 2008), or financial risk management (Baur et al., 2004) . In some jurisdictions such as the European Union, SA is officially required for policy assessment (Saltelli et al., 2019) . With roots in design of experiments (DOE), SA started to materialize in the 1970s and 1980s with the availability of computational resources and the extension of DOE to design of computer experiments (DACE); its large body of research is however spread across various disciplines, resulting in a lack of visibility (Razavi et al., 2021) . Why This Position Paper: ML e

FAQs

What is the main idea of the paper?

The paper argues that interpreting ML models and parts of the ML process can be viewed as a form of sensitivity analysis (SA). This enables a unified explanation framework and clarifies how related work should be credited.

How does sensitivity analysis relate to machine learning explanations?

SA studies how model outputs depend on model inputs. The paper connects ML interpretation efforts to SA by treating the ML process as a system suitable for SA and relating existing ML interpretation methods to that view.

Which ML components are explicitly linked to SA in the paper?

The paper connects interpretable ML and model-building with SA, including hyperparameter optimization. It also discusses relations among ML methods (e.g., partial dependence, FANOVA) and their SA roots, plus parallels to ablation analysis and one-factor-at-a-time approaches.

Aug 5, 2026

Upload by Genevieve