💻 computer science

How Class Imbalance Treatments Distort SHAP Attribution Fidelity: A Monte Carlo Investigation of Ranking and Magnitude Errors

Through a comprehensive Monte Carlo investigation, this paper demonstrates that while training sample size is the primary driver of SHAP attribution fidelity, common class imbalance treatments often distort feature rankings and magnitudes, leading the authors to recommend avoiding rebalancing in favor of class weighting when SHAP interpretability is a priority.

Original authors: Rıdvan Kara

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rıdvan Kara

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a crime. You have a computer assistant (the machine learning model) that looks at clues (data) to figure out who did it. To trust the computer, you ask it to explain why it pointed to a specific suspect. It gives you a list of reasons, saying, "Clue A was 40% responsible, Clue B was 10%," and so on. This explanation tool is called SHAP.

Now, imagine the crime is very rare. Maybe only 1 out of every 100 cases is a crime, while 99 are innocent people. This is called class imbalance. Because the "crime" examples are so scarce, the computer gets confused. To fix this, data scientists often try to "balance" the training data before teaching the computer. They might:

Copy-paste the rare crime examples (Random Oversampling).
Throw away most of the innocent examples (Random Undersampling).
Invent fake crime examples that look like real ones (SMOTE/ADASYN).
Tell the computer to pay extra attention to the rare crimes without changing the data (Class Weighting).

The big question this paper asks is: When we fix the data imbalance, do we accidentally break the computer's explanation? Does the list of "most important clues" stay the same? Do the numbers (like "40% responsible") stay accurate?

The author, Rıdvan Kara, ran a massive simulation (like running the same experiment 14,000 times with different settings) to find out. Here is what he discovered, using simple analogies:

1. The Size of the Class Matters More Than the Fix

The most important finding is that how many examples you have matters way more than which balancing trick you use.

The Analogy: Imagine trying to guess the flavor of a soup. If you only have a tiny spoonful (small sample size), it doesn't matter if you add salt, sugar, or pepper to the spoonful; your guess will be shaky. But if you have a giant pot of soup (large sample size), your guess will be accurate no matter what you did to the ingredients.
The Result: Increasing the amount of data you feed the computer is the single best way to get a reliable explanation. The choice of balancing method is about 3 to 5 times less important than simply having more data.

2. The "Ranking" vs. The "Magnitude" Trap

The paper makes a crucial distinction between two types of explanations:

Ranking: "Who is the top suspect?" (Is Clue A more important than Clue B?)
Magnitude: "How much did Clue A contribute?" (Did it contribute 10% or 50%?)

The study found that some balancing tricks are great at getting the order right but terrible at getting the numbers right.

The Analogy: Imagine a race.
- Random Oversampling (copy-pasting rare examples) is like a coach who makes sure the runner who should win is in first place (Ranking is good). However, the coach then tells everyone the winner ran twice as fast as they actually did (Magnitude is inflated/wrong).
- Random Undersampling (throwing away data) is like a coach who fires half the team. Now, they can't even tell who the best runner is, and the speed numbers are completely broken. They fail at both ranking and magnitude.

3. The Best "Do-Nothing" and "Soft" Fixes

When looking for the most honest explanation, the paper found two winners that didn't try to force the data to look balanced:

No Rebalancing: Just use the messy, unbalanced data as it is.
Class Weighting: Tell the computer, "Hey, the rare crimes are important, so pay extra attention to them," but don't delete or copy any data.

These two methods sat on the "Pareto frontier," meaning they were the best at both getting the order right and keeping the numbers accurate. They are the "minimally invasive" treatments.

4. The Synthetic "Fake" Data Problem

Methods that invent fake data (SMOTE and ADASYN) were okay at getting the order of suspects right, but they messed up the numbers.

The Analogy: It's like an artist who draws a fake face to fill a gap in a photo. The face looks like a person (the ranking is okay), but if you try to measure the distance between the eyes, the measurement is wrong because the face isn't real.

5. The "Extreme Rarity" Danger Zone

The paper found that these problems get much worse when the crime is extremely rare (less than 5% of cases) and you have very little data.

The Analogy: If you are trying to find a needle in a haystack, and you only have a tiny piece of hay to look at, any trick you use to "balance" the hay will likely make you miss the needle or guess the needle's size wrong. But if you have a whole barn of hay, you can find the needle easily, and the tricks matter less.

Summary of Recommendations

If you are using AI to explain decisions (like in medical diagnosis or credit scoring) and you have unbalanced data:

Get more data first. This is the most powerful tool.
Don't copy-paste or throw away data if you need accurate numbers. These methods distort the "how much" part of the explanation.
If you must balance, use Class Weighting (telling the model to care more) or just leave the data alone. These keep the explanation honest.
Check both the order and the numbers. Don't just look at which feature is #1; check if the percentage contribution makes sense, because some methods hide their mistakes in the numbers.

The paper concludes that while balancing data helps the computer predict better, it often breaks the computer's ability to explain itself accurately. If you need a trustworthy explanation, the gentlest approach (Class Weighting) or no approach at all is usually the best bet.

Technical Summary: How Class Imbalance Treatments Distort SHAP Attribution Fidelity

Problem Statement
SHAP (SHapley Additive exPlanations) is the standard for interpreting tree-based classifiers in high-stakes, class-imbalanced domains such as medical diagnosis, fraud detection, and credit scoring. Practitioners routinely apply class-imbalance treatments—random oversampling/undersampling, synthetic methods (SMOTE, ADASYN), and class weighting—to improve predictive performance. However, the implicit assumption that these upstream data manipulations do not distort downstream SHAP attributions remains untested. While the effects of these treatments on predictive metrics (AUC, F1) are well-documented, their impact on the fidelity of feature attributions is not. Specifically, it is unclear whether these treatments preserve the relative ranking of feature importance or the numerical magnitude of attributions, both of which are critical for different interpretability tasks (e.g., identifying top contributors vs. risk decomposition).

Methodology
The study employs a controlled Monte Carlo simulation framework to isolate the effects of imbalance treatments on SHAP fidelity, addressing three methodological gaps in prior literature:

Oracle Reference: Instead of using a class-balanced reference (which distorts magnitude scales), the study computes an oracle reference using near-infinite training data under the native class imbalance of each simulation cell.
Task Difficulty Calibration: The data-generating process (DGP) is calibrated so that the Bayes-optimal AUC remains constant (target 0.82 ± 0.01) across all cells. This ensures that observed differences in SHAP fidelity are due to the treatments themselves, not variations in baseline task difficulty.
Dual Fidelity Metrics: The study explicitly distinguishes between ranking fidelity (preservation of feature importance order) and magnitude fidelity (preservation of attribution values).

Experimental Design

Data Generation: Synthetic datasets with $p=20$ features drawn from a multivariate normal distribution. Two correlation structures were tested: independent features and block-correlated features (intra-block $\rho=0.6$ ).
Factors:
- Prevalence ( $\pi$ ): 0.02, 0.05, 0.10, 0.20.
- Sample Size ( $n$ ): 500, 2,000, 10,000.
- Correlation: Independent, Block.
- Treatments ( $k=6$ ): None (baseline), Random Oversampling, Random Undersampling, SMOTE, ADASYN, and Class Weighting.
Model: LightGBM with capacity-adaptive hyperparameters (specifically min_data_in_leaf scaled with post-balancing sample size to prevent model degeneration in extreme undersampling).
Scale: 24 design cells $\times$ 100 replications = 2,400 intended runs per treatment. Total successful fits: 14,371 (28 failures occurred in extreme rarity/small sample cells for synthetic methods due to nearest-neighbor constraints).
Evaluation: SHAP values were computed on a native-prevalence evaluation set (5,000 observations) to mirror deployment conditions. Fidelity was measured using 11 metrics: 6 rank-based (Kendall $\tau$ , Spearman $\rho$ , RBO, Jaccard@k, Recall@k) and 5 magnitude-based (L1, L2, Cosine similarity).

Key Results
The investigation yields four primary findings:

Sample Size Dominates Treatment Choice: Training sample size is the dominant driver of SHAP ranking fidelity, outweighing the choice of imbalance treatment by a factor of roughly 3 to 5. Increasing $n$ from 500 to 10,000 improves ranking fidelity (Kendall $\tau$ ) by ~0.22, whereas the spread between the best and worst treatments at a fixed $n$ is only ~0.03 to 0.07.
Extreme Rarity and Small Samples Amplify Gaps: The divergence between treatments is most pronounced under extreme rarity ( $\pi \le 0.05$ ) and small samples ( $n \le 2,000$ ). In these conditions, the treatment gap can rival the sample-size effect. Random undersampling performs particularly poorly here, often failing to produce meaningful models.
Divergence of Ranking and Magnitude Fidelity: The two axes of fidelity yield qualitatively different treatment rankings.
- Ranking: Random oversampling performs best (highest Kendall $\tau$ ), followed closely by SMOTE and "None." Random undersampling is consistently worst.
- Magnitude: "None" and Class Weighting jointly dominate, preserving the magnitude of attributions best. Random oversampling, while good for ranking, significantly inflates attribution magnitudes. Synthetic methods (SMOTE, ADASYN) underperform on magnitude. Random undersampling is the worst on both axes.
Pareto Frontier: "No rebalancing" and "Class weighting" jointly occupy the Pareto frontier, offering the best trade-off between preserving rank and magnitude. Random oversampling is only superior if the sole goal is ranking, but it distorts magnitudes. Random undersampling is dominated on both axes.

Significance and Claims
The paper claims that the assumption that imbalance treatments are neutral to SHAP attributions is false. The study establishes that:

Interpretability requires dual reporting: Relying solely on rank-based metrics conceals the magnitude inflation caused by oversampling and synthetic methods. Conversely, relying solely on magnitude metrics conceals the ranking advantages of oversampling. Both metric families must be reported to fully assess fidelity.
Default Recommendations: When SHAP attributions are a deliverable, practitioners should prefer minimally invasive treatments. "No rebalancing" and "Class weighting" are recommended as defaults because they preserve both the ordering and the magnitude of feature contributions.
Evaluation Distribution Matters: The study demonstrates that using a class-balanced evaluation set for SHAP analysis introduces magnitude distortions. A native-prevalence evaluation set is necessary to align SHAP magnitudes with real-world deployment scenarios.

The authors conclude that while predictive performance literature may justify certain resampling techniques, these choices carry a specific cost in interpretability fidelity. For applications where SHAP is used for risk decomposition, feature engineering, or quantitative reasoning, the distortion of magnitude by common resampling techniques (especially random oversampling and undersampling) poses a significant risk that can be mitigated by adopting class weighting or no rebalancing.