Navigating Shortcuts, Spurious Correlations, and Confounders: From Origins via Detection to Mitigation

TL;DR

This work identifies shortcuts, spurious correlations, and confounders as core failure modes of ML systems and offers a formal unification of these concepts via a comprehensive taxonomy. It links these phenomena to causality, bias, and security, and distinguishes world-induced from sampling-induced shortcuts, providing a formal framework grounded in distributions $P_ ext{gt}(x)$ and $P(x)$. The authors present a two-part taxonomy—detection and mitigation—supported by a structured view of available datasets, and they survey a wide range of methods: model-utility based, perturbation, XAI, and causality-driven detection, along with dataset-, model-, and inference-time mitigation strategies. The work culminates in a synthesis of open challenges, beyond-image-domain extensions, and a roadmap for standardized evaluation and dataset development to advance shortcut robustness in AI systems.

Abstract

Shortcuts, also described as Clever Hans behavior, spurious correlations, or confounders, present a significant challenge in machine learning and AI, critically affecting model generalization and robustness. Research in this area, however, remains fragmented across various terminologies, hindering the progress of the field as a whole. Consequently, we introduce a unifying taxonomy of shortcut learning by providing a formal definition of shortcuts and bridging the diverse terms used in the literature. In doing so, we further establish important connections between shortcuts and related fields, including bias, causality, and security, where parallels exist but are rarely discussed. Our taxonomy organizes existing approaches for shortcut detection and mitigation, providing a comprehensive overview of the current state of the field and revealing underexplored areas and open challenges. Moreover, we compile and classify datasets tailored to study shortcut learning. Altogether, this work provides a holistic perspective to deepen understanding and drive the development of more effective strategies for addressing shortcuts in machine learning.

Figures (7)

• Figure 1: Models across different settings are susceptible to shortcuts. Models trained on data containing spurious correlations may rely on unintended features for decision-making. These shortcuts can manifest across various domains and tasks, significantly affecting model performance and generalization.
• Figure 2: Overview where spurious correlations can appear. Given our established example of classifying birds into landbirds and waterbirds (based on their characteristics), the environment is a spurious feature naturally occurring in the world (i). The distorted ($\rightsquigarrow$) sampling process can then induce spurious correlations through, for example, photographer tags (ii).
• Figure 3: Overview of the waterbirds example in the context of causality. In the data, we have access to the bird's characteristics and its environment and we want to predict whether the bird is a landbird or waterbird. If we assume that the bird's characteristics (i.e., its appearance and abilities) cause both its environment and whether it is a waterbird, environment and the target label are correlated in the data (while not causally related). The common cause (bird characteristics) is called a confounder.
• Figure 4: Overview of our taxonomy on shortcut learning. We categorize approaches into the two areas of shortcut detection and shortcut mitigation. Detailed information, including all subcategories, is provided in \ref{['fig:detection']} and \ref{['fig:mitigation']} in the following sections.
• Figure 5: Taxonomy of shortcut detection approaches. A comprehensive breakdown of shortcut detection methods, organized into methodological subcategories.