🤖 machine learning

Correcting Performance Estimation Bias in Imbalanced Classification with Minority Subconcepts

This paper introduces predicted-weighted balanced accuracy (pBA), a practical evaluation metric that mitigates performance estimation bias in imbalanced classification by replacing unavailable true subconcept labels with predicted posterior probabilities to provide more stable and interpretable assessments of model performance across heterogeneous subpopulations.

Original authors: Taylor Maxson, Roberto Corizzo, Yaning Wu, Nathalie Japkowicz, Colin Bellinger

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Taylor Maxson, Roberto Corizzo, Yaning Wu, Nathalie Japkowicz, Colin Bellinger

Original paper licensed under CC BY 4.0 (http://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

The Big Problem: The "Average" Lie

Imagine you are a principal evaluating a new school lunch program. You ask, "How did the students like the food?" The principal says, "Great! The average rating is 9 out of 10."

Sounds good, right? But what if the "average" is hiding a secret?

Group A (The Majority): 90% of the students are eating standard sandwiches. They love them (10/10).
Group B (The Minority): 10% of the students have severe nut allergies and are eating a special, bland, gluten-free meal. They hate it (1/10).

If you just look at the average, the program looks like a huge success (9/10). But if you look at the subgroups, you see that the program is actually a disaster for the group with the allergies. The "average" score is lying to you because it lets the huge group of sandwich-eaters drown out the small group of allergy sufferers.

This is exactly what happens in Machine Learning when dealing with Imbalanced Data.

The Class: "Sick Patients" (Minority) vs. "Healthy Patients" (Majority).
The Subconcepts: Within the "Sick" group, there might be "Flu" (common) and "Rare Genetic Disorder" (very rare).

If a doctor's AI model is 95% accurate on "Flu" but only 10% accurate on the "Rare Genetic Disorder," the overall score might still look great. But in the real world, that AI is failing the people who need help the most. The paper calls this Performance Estimation Bias.

The Old Way vs. The New Way

The Old Way (Unweighted Scores):
This is like the school principal just taking the average. It assumes every student counts the same, regardless of how many there are. In data terms, if a "Rare Disease" makes up only 1% of your test data, the computer barely notices if it gets them wrong. The score stays high, giving a false sense of security.

The Previous "Fix" (True Weights):
Researchers previously tried to fix this by saying, "Let's count the Rare Disease patients 100 times more than the Flu patients." This works perfectly, BUT it requires a superpower: you need to know exactly which specific disease a patient has before you test the model. In the real world, you usually don't know this until after the fact. It's like trying to grade the lunch program by knowing every student's allergy before they even take a bite.

The Paper's Solution (Predicted-Weighted Balanced Accuracy or "pBA"):
The authors invented a clever workaround. They realized you don't need to know the sub-concept for sure; you just need a good guess.

The Training Phase: They teach a helper AI (a "Subconcept Classifier") to recognize the different types of "Sick" patients (Flu vs. Rare Disorder) using data where the labels are known.
The Testing Phase: When a new patient comes in, the main AI makes a diagnosis. Simultaneously, the helper AI looks at the patient and says, "I'm 80% sure this is Flu, but 20% sure it's the Rare Disorder."
The Magic Math: Instead of forcing a hard choice (Flu OR Disorder), the new method uses that 80/20 guess to adjust the score.
- If the main AI gets the "Flu" right, it gets a normal score.
- If the main AI gets the "Rare Disorder" wrong, but the helper AI was uncertain (maybe it thought it was Flu), the penalty is softer.
- If the main AI gets the "Rare Disorder" wrong and the helper AI was confident it was a Rare Disorder, the penalty is heavy.

This creates a "Soft, Uncertainty-Aware" score. It doesn't just look at the final answer; it looks at how confident the system was in the type of patient it was treating.

Why This Matters (The "Why Should I Care?")

The paper tested this on three types of real-world data:

Tabular Data: Like spreadsheets of numbers (e.g., credit scores, vehicle types).
Medical Imaging: X-rays of lungs (looking for different types of lung issues).
Text: Detecting hate speech (looking for different types of hate speech).

The Results:

The "Average" Score is often a liar. In many cases, the standard score was very high, but the model was actually failing the small, rare groups.
The New Score (pBA) tells the truth. It drops the score when the model is failing the rare groups, even if the test set is mostly full of common cases.
It's not about making the score lower. Sometimes, if the rare groups are actually easier to predict than the common ones, the new score goes up. It's not trying to be pessimistic; it's trying to be accurate about where the model is actually good or bad.

The Takeaway Analogy

Imagine you are judging a Talent Show.

The Old Score: You count every vote. If 900 people vote for the "Singer" and 10 people vote for the "Juggler," the Singer wins 99% of the time. You never know if the Juggler is actually terrible.
The Paper's Score: You realize the Juggler is a "rare act." You ask the audience, "How sure are you this is a Juggler?"
- If the audience is confused (uncertain), you don't penalize the Juggler too hard for a mistake.
- If the audience is sure it's a Juggler, and the Juggler fails, you give them a big "F."
- This gives you a Fairer Report Card that tells you: "The Singer is great, but the Juggler needs more practice," even though the Juggler only had 10 votes.

Summary

This paper introduces a new way to grade AI models that doesn't let the "popular" groups hide the failures of the "rare" groups. It uses a "guessing" system to adjust the grades, ensuring that if an AI fails a small, important group, the final score reflects that failure, preventing dangerous mistakes in fields like medicine or safety.

1. Problem Statement

The paper addresses a critical flaw in evaluating imbalanced classification models: performance estimation bias caused by within-class heterogeneity.

The Core Issue: Standard evaluation metrics (e.g., Balanced Accuracy, F1, AUC) operate at the class level. However, a single class (especially a minority class) often contains multiple distinct subconcepts (e.g., different disease subtypes in medical imaging or different hate-speech categories in text).
The Bias: In imbalanced settings, standard metrics are heavily biased toward the largest minority subconcepts. A model can achieve a high aggregate score by performing well on the most frequent subconcept while failing catastrophically on smaller, rarer, but clinically or socially critical subconcepts.
The Limitation of Prior Work: Previous research showed that using true subconcept labels to weight instances during evaluation could correct this bias. However, in real-world deployment, true subconcept labels are rarely available at test time, making this correction impractical.

2. Methodology

The authors propose a practical, utility-weighted evaluation framework that approximates the ideal "true-label" weighting using only training data and predicted probabilities.

A. Conceptual Framework: Utility-Weighted Balanced Accuracy (UWBA)

The authors redefine Balanced Accuracy (BA) to include instance weights ( $w_i$ ) that reflect the relative importance of subconcepts.

Standard BA: $BA = \frac{1}{2} (Recall_0 + Recall_1)$ .
Weighted Recall: $Recall^{(w)}_c = \frac{\sum w_i \cdot \mathbb{I}[y_i=c \land \hat{y}_i=c]}{\sum w_i \cdot \mathbb{I}[y_i=c]}$ .
Goal: Assign higher weights to instances belonging to smaller subconcepts to counteract the dominance of large subconcepts in the aggregate score.

B. The Proposed Solution: Predicted-Weighted Balanced Accuracy (pBA)

Since true subconcept labels are unavailable at test time, the authors introduce a soft, uncertainty-aware weighting strategy:

Training Phase: A multiclass subconcept classifier is trained on the minority class (and potentially the majority class) using available subconcept labels.
Inference Phase: For a test instance $x$ , the subconcept classifier outputs a posterior probability distribution $q_s(x) = P(s|x)$ over all subconcepts $s$ .
Weight Calculation: Instead of hard-assigning an instance to a single subconcept (which introduces brittle errors), the weight is calculated as the expected utility:
$a_x = \sum_{s \in S} q_s(x) u_s$
Where $u_s$ is a utility value assigned to subconcept $s$ (typically inversely proportional to its size in the training set).
Result: This yields pBA (Predicted-Weighted Balanced Accuracy), a metric that is sensitive to performance on rare subconcepts without requiring ground-truth subconcept labels at test time.

3. Key Contributions

Practical Evaluation Metric: The introduction of pBA, which bridges the gap between theoretical ideal evaluation (using true labels) and practical deployment (where labels are missing).
Soft Weighting Mechanism: Moving away from hard assignment of subconcepts to a probabilistic, expectation-based weighting that preserves uncertainty and avoids penalizing the metric for classifier boundary errors.
Comprehensive Empirical Analysis: The study validates the method across diverse domains:
- Tabular Benchmarks: Keel and PMLB datasets.
- Medical Imaging: NIH ChestX-ray, Ottawa Hospital, Pneumonia/COVID-19/TB, and HAM10000.
- Text/Multimodal: MMHS150K (Hate Speech).
Diagnostic Insight: The paper demonstrates that pBA acts as a diagnostic tool. When pBA diverges significantly from standard unweighted BA, it signals that the aggregate score is misleading and driven by specific subpopulations.

4. Experimental Results

The authors evaluated the method against three Research Questions (RQs):

RQ1: Can predicted weighting reduce bias toward large subconcepts?
- Result: Yes. On Keel and PMLB datasets, pBA significantly reduced the correlation gap between the full-test score and the largest minority subconcept score.
- Comparison: While "True-Weighted BA" (using ground truth) provided the strongest correction, pBA successfully approximated this behavior, moving the aggregate score away from being dominated by the largest subconcept.
- Metric Sensitivity: The effect was most pronounced for Balanced Accuracy (BA) and F1-score. AUC showed less dramatic improvement, likely because it is a ranking-based metric less sensitive to specific instance weighting than accuracy-based metrics.
RQ2: How does subconcept classifier accuracy affect pBA reliability?
- Result: There is a strong negative correlation between subconcept classifier performance and the error between predicted and true weights.
- Implication: As the subconcept classifier improves, the pBA weights converge toward the ideal true weights. This confirms that the method is robust and improves as the auxiliary model improves.
RQ3: How do subconcept geometry and difficulty affect the scores?
- Result: The method does not blindly penalize rare subconcepts.
  - If small subconcepts are harder (lower accuracy), pBA decreases the aggregate score, exposing the optimism of the unweighted metric.
  - If small subconcepts are easier (higher accuracy), pBA can increase the aggregate score.
- Significance: This proves pBA is not a "pessimistic" correction but a representative one, reflecting the actual distribution of performance across the latent subpopulations.

5. Significance and Conclusion

Beyond Fairness: The paper frames this not just as a fairness issue, but as a fundamental evaluation problem. High average performance can mask "hidden stratification" where critical subgroups are underserved.
Deployment Safety: In high-stakes domains like medicine, relying on unweighted scores can lead to deploying systems that fail on rare but critical conditions. pBA provides a more stable and interpretable assessment.
Lightweight Diagnostic: The method does not require retraining the main classifier or complex group-robust optimization. It serves as a lightweight post-hoc layer to determine if a reported score is robust to within-class heterogeneity.
Future Work: The authors suggest improving the subconcept prediction step and extending the analysis to large majority subconcepts.

Summary: The paper introduces pBA, a practical evaluation metric that uses predicted subconcept probabilities to correct the bias of standard imbalanced classification metrics. It effectively reveals when a model's high average score is an illusion caused by the dominance of large, easy subconcepts, offering a more reliable tool for assessing model performance in heterogeneous, real-world scenarios.