On the Need for a Language Describing Distribution Shifts: Illustrations on Tabular Datasets

Jiashuo Liu, Tianyu Wang, Peng Cui, Hongseok Namkoong

Introduction

Machine learning models are notoriously fragile; they often degrade when the "test" environment differs from the "training" environment—a phenomenon known as distribution shift. Historically, the research community has focused heavily on covariate shift ($X$-shift). In these scenarios, the distribution of features changes (e.g., a self-driving car encountering snow for the first time), but the underlying "rules" mapping inputs to outputs remain the same.

However, in tabular data—such as healthcare records, census data, or financial logs—the rules themselves often change due to hidden confounders. For example, a model predicting income based on education might perform well in California but fail in Puerto Rico because it lacks context on local language requirements or differing economic structures. This is known as a conditional shift ($Y|X$-shift).

This paper argues that our current "language" for describing distribution shifts is too coarse, leading researchers to apply the wrong algorithmic fixes to the wrong problems. To address this, the authors introduce WHYSHIFT, a massive benchmark proving that conditional shifts break common performance assumptions and require targeted, data-centric interventions.

Relative regret across standard benchmarks versus the proposed WHYSHIFT datasets, highlighting the severity of natural shifts.

The WHYSHIFT Benchmark and Methodology

To move beyond simulated shifts, the authors conducted a massive empirical study using real-world data and proposed a new diagnostic framework:

  1. The WHYSHIFT Benchmark: The authors curated five real-world tabular datasets (including US Census data, US traffic accidents, and international taxi ride durations) and constructed 22 different "settings" based on natural geographic and temporal shifts. They evaluated 22 different algorithms—ranging from standard Gradient Boosted Trees to complex Distributionally Robust Optimization (DRO) methods—across over 86,000 hyperparameter configurations.
  2. Shift Decomposition: Using a framework called DISDE, they mathematically separated how much of a model's performance drop was due to $X$-shift versus $Y|X$-shift.
  3. Risk Region Identification: They developed a method to find specific "regions" in the data where the $Y|X$-shift is most severe. By training a domain classifier and reweighting samples, they can identify which specific subpopulations (e.g., "college-educated managers") experience the most drastic changes in feature-label relationships.

Key Findings

$Y|X$-shifts are the Rule, Not the Exception

In tabular data, $Y|X$-shifts are far more prevalent than in domains like computer vision. Out of 169 source-target pairs exhibiting significant performance degradation in their benchmark, 87.2% had over half of that drop attributed to $Y|X$-shifts.

The "Accuracy-on-the-Line" Myth Fails

In vision tasks, a model’s out-of-distribution (OOD) performance usually correlates linearly with its in-distribution performance—a trend known as "accuracy-on-the-line." This paper demonstrates that this trend completely breaks down when $Y|X$-shifts are strong. Being highly accurate on your training data offers no guarantee of success on the target data if the underlying relationship has fundamentally changed.

Target vs. Source accuracy across various datasets. High $Y|X$-shift breaks the linear correlation, while low $Y|X$-shift maintains it.

Algorithms Aren't a Silver Bullet

Popular "robust" methods like DRO are incredibly sensitive to tuning and often fail to meaningfully improve robustness. Despite the hype around advanced optimization techniques, standard tree ensembles (like XGBoost and LightGBM) remain the most competitive models on tabular data, even though they still suffer under severe shifts.

The Power of Missing Features and Targeted Data

$Y|X$-shifts are frequently just "missing $X$-shifts." In a case study predicting income, the authors found that adding a single missing feature—English language proficiency—collapsed a massive $Y|X$-shift into a manageable one.

Furthermore, data-centric interventions proved vastly superior to algorithmic ones. By collecting just 250 targeted samples from a "risk region" identified by their algorithm, the authors improved OOD accuracy more effectively than by collecting random data or applying complex algorithmic fixes.

Limitations and Future Directions

While WHYSHIFT provides a crucial step forward, the authors note some limitations. The benchmark currently focuses on economic and transportation data; medical or biological tabular data might exhibit different shift patterns. Additionally, their region-identification algorithm requires some labeled data from the target domain, which isn't always available in blind deployment scenarios. Finally, they caution that targeting data collection on specific "risk regions" must be done carefully to avoid privacy or fairness concerns regarding marginalized groups.

Ultimately, the paper calls for a paradigm shift: moving away from "one-size-fits-all" robust algorithms and toward diagnostic tools that help practitioners understand why a distribution changed, enabling smarter, data-driven interventions.

Summary updated

Abstract

Different distribution shifts require different algorithmic and operational interventions. Methodological research must be grounded by the specific shifts they address. Although nascent benchmarks provide a promising empirical foundation, they \emph{implicitly} focus on covariate shifts, and the validity of empirical findings depends on the type of shift, e.g., previous observations on algorithmic performance can fail to be valid when the $Y|X$ distribution changes. We conduct a thorough investigation of natural shifts in 5 tabular datasets over 86,000 model configurations, and find that $Y|X$-shifts are most prevalent. To encourage researchers to develop a refined language for distribution shifts, we build ``WhyShift``, an empirical testbed of curated real-world shifts where we characterize the type of shift we benchmark performance over. Since $Y|X$-shifts are prevalent in tabular settings, we \emph{identify covariate regions} that suffer the biggest $Y|X$-shifts and discuss implications for algorithmic and data-based interventions. Our testbed highlights the importance of future research that builds an understanding of why distributions differ.
BibTeX
@inproceedings{liu2023need,
  author = {Jiashuo Liu and Tianyu Wang and Peng Cui and Hongseok Namkoong},
  title = {{On the Need for a Language Describing Distribution Shifts: Illustrations on Tabular Datasets}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=PF0lxayYST},
}
Figure