The Pragmatic Frames of Spurious Correlations in Machine Learning: Interpreting How and Why They Matter

Technical Summary: The Pragmatic Frames of Spurious Correlations in Machine Learning

Problem Statement
Machine learning (ML) relies on learning correlations from data to make predictions. However, models frequently capture "spurious correlations"—relationships that are statistically present in training data but do not reflect the underlying causal mechanisms of the target phenomenon. While the statistical definition of spuriousness is clear (a non-causal relationship arising from coincidence or confounding), the ML literature reveals a disconnect between this formal definition and how researchers actually identify and address the problem in practice.

Standard ML pipelines often lack the capacity to distinguish causation from correlation. Consequently, when models fail due to unintended correlations (e.g., classifying cows based on grass backgrounds rather than animal features), researchers must determine why the correlation is problematic. The paper argues that researchers rarely rely on formal causal definitions to label a correlation as spurious. Instead, they engage in "sense-making" through alternative lenses that prioritize technical performance, task alignment, or ethical outcomes over strict causal inference.

Methodology
The authors conducted a two-stage critical literature analysis to identify how spuriousness is framed and justified in ML research.

1. Data Collection: Using the Semantic Scholar API, the authors retrieved 713 papers containing the terms "spurious correlation," "spurious feature," or "shortcut" from ML, NLP, and computer vision venues (including preprints, journal articles, and conference papers).
2. Sampling and Coding:
   • Stage 1: A subset of 200 papers was randomly sampled for manual screening, resulting in 49 relevant papers. Through bibliography review, 16 additional high-impact works were added, creating an initial analytical sample of 65 papers. Two authors independently coded these papers to extract recurring rationales for labeling correlations as spurious.
   • Stage 2: The established coding framework was applied to the remaining 513 papers (after filtering for relevance and quality), yielding an additional 222 articles.
3. Final Sample: The analysis culminated in a total sample of 287 papers. The authors identified four dominant "pragmatic frames" used to interpret spuriousness, noting that papers often invoked multiple frames simultaneously.

Key Contributions: The Four Pragmatic Frames
The paper posits that the desirability of a correlation is not a fixed statistical property but a situated judgment informed by "pragmatic frames." These frames represent specific ways of seeing that help researchers isolate and interpret aspects of the problem based on function and normative goals rather than fixed definitions. The four identified frames are:

1. Relevance:

   • Definition: A correlation is spurious if it is not relevant to the specific learning task, even if it is predictive in the training data.
   • Rationale: Researchers expect models to utilize features that are inherently germane to the target label (e.g., the shape of an object) rather than incidental artifacts (e.g., background color or texture).
   • Critique: "Relevance" is often defined by the constraints of the dataset and the researcher's intent (e.g., assuming background is irrelevant in single-label image classification) rather than the intrinsic properties of the data.

2. Generalizability:

   • Definition: A correlation is spurious if it fails to hold across different data distributions, specifically when moving from training to test data or under "out-of-distribution" (OOD) shifts.
   • Rationale: Models should learn stable, environment-invariant relationships. Correlations that are "environment-specific" or rely on "misleading heuristics" that do not generalize are deemed spurious.
   • Critique: Researchers rarely specify which distributions a model must generalize to. The definition of a "stress test" or "natural distribution shift" is contingent on researcher expectations and the practical ability to construct representative test sets.

3. Human-likeness:

   • Definition: A correlation is spurious if it relies on decision rules or features that a human would not use to perform the same task.
   • Rationale: This frame assumes a "human-like" function is the gold standard. For instance, if humans recognize objects by shape but a model relies on texture, the model's correlation is considered spurious.
   • Critique: This frame presupposes a universal, stable account of human cognition, which is contested by evidence of cultural and demographic variation in human perception. Furthermore, it may conflict with goals to build systems that surpass human performance or operate differently to avoid human biases.

4. Harmfulness:

   • Definition: A correlation is spurious if its use leads to socially or ethically harmful outcomes, such as amplifying biases against protected groups.
   • Rationale: Even if a correlation is statistically stable and predictive, it is deemed undesirable if it encodes societal stereotypes (e.g., associating "CEO" with men) or leads to performance disparities for minority groups.
   • Critique: Determining "harm" requires normative judgments about how the world should be represented versus how it is. A correlation reflecting a stable fact (e.g., gender distribution in nursing) may be labeled spurious if it reinforces stereotypes, highlighting that the frame is driven by ethical goals rather than data properties.

Results and Findings

• Prevalence: Among the 287 papers analyzed, Generalizability was the most frequently invoked frame (44%), followed by Relevance (36%), Harmfulness (28%), and Human-likeness (18%).
• Operationalization: The study found that researchers rarely define spuriousness using formal causal theory. Instead, they use these pragmatic frames as motivations for developing solutions (e.g., robustness techniques, fairness constraints).
• Contingency: The classification of a correlation as "spurious" is highly contingent on the researcher's intent, the specific task, and the normative goals of the project. What is considered a "shortcut" in one context (e.g., background color in object recognition) might be a valid feature in another.

Significance and Claims
The paper claims to contribute to broader conversations on the epistemic culture of machine learning by demonstrating that technical concepts like "spuriousness" are defined and operationalized through contingent practices rather than static definitions.

• Epistemic Shift: The authors argue that by relying on pragmatic frames, researchers transform spuriousness from a static property of data into a dynamic relationship between data and researcher expectations.
• Normative Implications: The choice of which frame to prioritize (e.g., prioritizing relevance over human-likeness) is a value-laden decision with trade-offs. The paper suggests that the ML community often obscures these normative choices by using the term "spurious" as a technical shorthand.
• Future Directions: The authors propose that acknowledging these frames can serve as a resource for improving ML methods. Rather than ignoring the pragmatic nature of these judgments, future work could develop benchmarks and methods that explicitly address the specific frames researchers care about (e.g., designing for specific distribution shifts or explicitly reducing harm).
• Modesty: The authors explicitly state that their typology is not exhaustive and that as ML research evolves into multimodal and agent-based domains, new framings may emerge. They emphasize that their goal is to offer a theoretical lens for interrogating the limits of correlation-based ML, not to provide a definitive causal solution.