An experimental study on the Rashomon effect of balancing methods in imbalanced classification

Mustafa Cavus, Przemyslaw Biecek

TL;DR

While data balancing techniques like SMOTE are essential for handling imbalanced datasets, they significantly inflate the Rashomon effect—a phenomenon where multiple models achieve near-identical accuracy but produce conflicting predictions for the same individuals. This paper introduces a new metric, obscurity, to quantify this "predictive multiplicity" and warns that balancing data without monitoring these metrics can lead to "blind" and potentially irresponsible model selection.

What Problem Does This Paper Address?

In machine learning, we often encounter "imbalanced" datasets where one class (e.g., credit card fraud) is much rarer than another (e.g., legitimate transactions). Standard models tend to ignore the minority class to maximize overall accuracy. To fix this, researchers use "data-centric AI" techniques like oversampling the minority or undersampling the majority.

However, the authors point out a hidden danger: the Rashomon Effect. Named after the classic Kurosawa film where witnesses give contradictory accounts of the same event, this effect occurs when many different models perform equally well on a dataset but disagree on specific predictions. If you have ten "best" models and they all give different answers for the same patient or applicant, picking just one is a gamble. This paper investigates how common balancing methods impact this uncertainty.

Methods

The researchers utilized a tree-based AutoML tool called forester to generate a Rashomon Set—a collection of models whose performance is within a small margin ($\epsilon = 0.05$) of the best-performing "reference" model.

To measure the disagreement within these sets, they used three key metrics:

  1. Ambiguity: The percentage of samples where at least one model in the set disagrees with the reference model.
  2. Discrepancy: The maximum possible disagreement between the reference model and any other single model in the set.
  3. Obscurity (New): A novel metric proposed by the authors that calculates the average ratio of conflicting predictions across all samples and all models in the set. This provides a more nuanced, data-centric view of how "confusing" the Rashomon set actually is.

They also introduced Variable Importance Order Discrepancy (VIOD), which uses Kendall’s $\tau$ correlation to see if models in the Rashomon set prioritize the same features.

Datasets

The study evaluated 21 imbalanced benchmark tabular datasets from various domains (e.g., spambase, wine_quality, mammography). The imbalance ratios in these datasets ranged from mild (1.54:1) to extreme (129.53:1).

Key Insights and Ideas

  • Balancing is a Double-Edged Sword: While balancing methods (like SMOTE or Random Oversampling) successfully improve performance on the minority class, they systematically increase the "obscurity" and "discrepancy" of the models. In short: fixing the bias makes the models more prone to disagreement.
  • Obscurity vs. Ambiguity: The authors argue that "Ambiguity" is too blunt because it only tells you if a conflict exists. "Obscurity" reveals the density of that conflict, making it a better tool for researchers to understand the stability of their predictions.
  • Partial Resampling isn't a Cure: Some researchers suggest "partial resampling" (only balancing the data slightly, e.g., to a 1.25:1 ratio) to avoid overfitting. However, this paper found that partial resampling does not significantly reduce predictive multiplicity compared to full balancing.
  • Feature Importance is Stable: Interestingly, while the models disagreed on specific predictions, they generally agreed on which features were important (low VIOD). This suggests the Rashomon effect in these cases is driven by how models handle specific data points rather than a fundamental disagreement on the underlying logic.

Key Results

  • Multiplicity Inflation: Statistical tests (Kruskal-Wallis) confirmed that all balancing methods significantly increased obscurity and discrepancy compared to models trained on the original, imbalanced data.
  • Oversampling vs. Undersampling: Oversampling methods (Random Oversampling and SMOTE) generally provided better performance gains than undersampling. However, they also led to Rashomon sets with higher predictive multiplicity.
  • The Trade-off Plot: The authors demonstrated an "extended performance-gain plot." By plotting AUC gain on the x-axis and Obscurity/Discrepancy on the y-axis, practitioners can visually identify the "sweet spot" where they get the best performance boost with the least amount of added predictive instability.

Limitations and Future Work

The authors acknowledge a few areas for further exploration:

  • Model Scope: The study focused primarily on tree-based models via the forester tool. Future work could examine if deep learning or linear models exhibit the same inflation of multiplicity when balanced.
  • Data Complexity: There is a hypothesis that oversampling increases "data complexity," which in turn fuels the Rashomon effect. The authors suggest this link deserves a dedicated study.
  • Cost-Sensitive Learning: The paper focused on data-centric balancing. It remains to be seen if model-centric approaches (like changing the loss function to penalize minority class errors) produce less predictive multiplicity than resampling the data.

Summary updated

Abstract

Predictive models may generate biased predictions when classifying imbalanced datasets. This happens when the model favors the majority class, leading to low performance in accurately predicting the minority class. To address this issue, balancing or resampling methods are critical data-centric AI approaches in the modeling process to improve prediction performance. However, there have been debates and questions about the functionality of these methods in recent years. In particular, many candidate models may exhibit very similar predictive performance, called the Rashomon effect, in model selection, and they may even produce different predictions for the same observations. Selecting one of these models without considering the predictive multiplicity -- which is the case of yielding conflicting models' predictions for any sample -- can result in blind selection. In this paper, the impact of balancing methods on predictive multiplicity is examined using the Rashomon effect. It is crucial because the blind model selection in data-centric AI is risky from a set of approximately equally accurate models. This may lead to severe problems in model selection, validation, and explanation. To tackle this matter, we conducted real dataset experiments to observe the impact of balancing methods on predictive multiplicity through the Rashomon effect by using a newly proposed metric obscurity in addition to the existing ones: ambiguity and discrepancy. Our findings showed that balancing methods inflate the predictive multiplicity and yield varying results. To monitor the trade-off between the prediction performance and predictive multiplicity for conducting the modeling process responsibly, we proposed using the extended version of the performance-gain plot when balancing the training data.
BibTeX
@misc{cavus2024experimental,
  author = {Mustafa Cavus and Przemyslaw Biecek},
  title = {{An experimental study on the Rashomon effect of balancing methods in imbalanced classification}},
  year = {2024},
  eprint = {2405.01557},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.01557},
}
Figure