Fraud, Cancer, and the Fall of SMOTE: Class Imbalance in Practice and the 2026 Shift to Cost-Sensitive Learning

Fraud detection and cancer screening share one statistical problem: the thing you care about is rare. A fraction of a percent of card transactions are fraudulent. Most scans come back clean. Since SMOTE arrived in the early 2000s, the standard fix was to manufacture more rare cases synthetically. That fix is losing its default status, and the teams with the most at stake are leading the retreat.

Resampling Just Lost Its Default Status

Thesis: The structural shift in handling Class Imbalance isn’t a new algorithm. It’s the demotion of blanket resampling from automatic default to one option among several, with cost-sensitive methods taking the front seat.

For most of two decades the playbook was simple. Skewed classes break naive accuracy, so you rebalance the data first: synthesize minority rows until the classes roughly match, then train. SMOTE, the Synthetic Minority Over-sampling Technique from Chawla and colleagues, was the workhorse, and Oversampling became a default step, not a decision.

The reappraisal flips that order. The new instinct is to leave the data alone and change the model’s incentives instead: tell the algorithm that missing a rare fraud case costs far more than a false alarm. That is Cost Sensitive Learning, and it’s moving from niche technique to first move.

This isn’t SMOTE dying. It’s SMOTE getting demoted.

The Evidence Has Been Stacking for Years

Two studies did most of the damage, and they didn’t land on the same conclusion by accident.

In 2022, an Amazon team ran SMOTE and its variants against strong classifiers across dozens of datasets. Once the classifier is strong, they found, the benefit from balancing shrinks toward zero, and tuning the decision threshold often matches resampling outright (Elor’s SMOTE study).

Two years later, a separate group went further. Across most of the datasets they tested, no rebalancing at all was competitive with a tuned LightGBM, random forest, or even plain logistic regression (Sakho et al. rebalancing study). They also showed the mechanism: with default settings, SMOTE essentially duplicates minority samples, and its synthetic points thin out near the class boundary, exactly where the model needs signal most.

Different teams. Different datasets. Same verdict. That’s not noise, that’s a pattern.

The field agrees. In credit-card fraud, systems now lean on cost-sensitive learning and threshold optimization, sometimes combined with resampling rather than relying on it alone (Frontiers in AI). In rare-disease and cancer detection, teams use hybrid resampling, cost-sensitive losses, and focal-loss variants, sometimes reframing the rare class as anomaly detection (Springer AI Review). The reflex moved from “balance the data” to “price the errors.”

Who Wins the Reappraisal

The winners share one trait: they treat imbalance as a modeling problem, not a data problem.

Cost-sensitive learning is the clearest beneficiary. It’s already built into the tools through Class Weighting via the class_weight argument in scikit-learn and Keras, and scale_pos_weight in XGBoost and LightGBM. No synthetic data, no extra pipeline stage, no leakage surface. One parameter, and the model starts paying attention to the rare class (imbalanced-learn on PyPI).

Threshold tuning wins too. Instead of forcing a 50% cutoff on a skewed problem, you move the decision boundary to where precision and recall balance for your cost structure. Cheap, transparent, and paired with proper metrics: PR-AUC and Balanced Accuracy, not raw accuracy, which lies on skewed data.

Deep learning teams get Focal Loss, which down-weights the easy, abundant examples so training spends its budget on the rare ones.

The Imbalanced Learn library isn’t a casualty either. It’s still the standard toolkit; it just stopped being a one-click SMOTE button and started teaching evaluation.

In 2026, you’re either modeling the imbalance or you’re manufacturing data to hide it.

Who’s Still Running the Old Playbook

The losers aren’t using bad tools. They’re using the right tools for a problem the field already moved past.

Anyone reaching for SMOTE as an automatic first step is optimizing last decade’s workflow. The technique still has real uses, but as a reflex it’s a liability, especially when teams bolt it on without measuring whether it beats doing nothing.

The sharper risk is Data Leakage. Run resampling before the train/test split, a stunningly common mistake in tutorials, and synthetic minority rows bleed across the boundary. Your validation score looks great. Production doesn’t.

Undersampling the majority class has the opposite failure: throw away real data to force balance, and you starve the model of signal. The SMOTE successors, including ADASYN, Borderline-SMOTE, and the hybrids, carry mixed evidence, not a clean win.

The teaching material is the real loser. Every bootcamp still presenting “import SMOTE, rebalance, done” as best practice is shipping graduates into a consensus that already moved. You’re either updating the curriculum or training people for last year’s job.

What Happens Next

Base case (most likely): Cost-sensitive learning, threshold tuning, and proper metrics become the documented default in libraries and courses. SMOTE survives where it earns its place, inside hybrid pipelines for severe imbalance, but stops being the reflex. Signal to watch: New tutorials and library docs leading with class_weight and PR-AUC instead of SMOTE. Timeline: Already underway; mainstream within a year or two.

Bull case: Tooling makes cost-sensitive defaults automatic, evaluation on balanced accuracy and PR-AUC becomes standard, and leakage-by-resampling drops sharply. Signal: Default model configs shipping with sane class weights and threshold calibration. Timeline: Two to three years for broad adoption.

Bear case: The nuance flattens into a new dogma, “SMOTE is dead,” and teams swing to the opposite extreme, dropping resampling even where hybrids genuinely help in fraud and rare-disease work. Signal: Blog posts declaring SMOTE obsolete with no dataset caveats. Timeline: The risk is live now.

The shift is real. The only open question is whether the field overcorrects.

Frequently Asked Questions

Q: How do banks handle class imbalance in fraud detection systems? A: Modern fraud systems lean on cost-sensitive learning and decision-threshold tuning, penalizing missed fraud more heavily than false alarms, often combined with resampling rather than replacing it. Evaluation uses PR-AUC and recall at fixed precision, never raw accuracy, since fraud is a fraction of a percent of transactions.

Q: How do medical AI models detect rare diseases with imbalanced data? A: Rare-disease and cancer models combine hybrid resampling, cost-sensitive losses, and focal-loss variants that down-weight common cases. Some reframe detection as anomaly detection, treating the rare condition as an outlier. The constant is evaluation with balanced accuracy and PR-AUC, because raw accuracy rewards a model that simply ignores the rare class.

Q: Why are practitioners moving away from SMOTE to cost-sensitive learning in 2026? A: Because peer-reviewed studies keep showing SMOTE often fails to beat no rebalancing once you use modern gradient boosting, while adding leakage and calibration risk. Cost-sensitive learning reaches the same goal with one parameter, no synthetic data, and no extra pipeline stage where mistakes can hide.

The Bottom Line

Class imbalance didn’t get easier. The field just got honest about what actually works. The reflex to synthesize your way out of a skewed dataset is giving way to a cheaper, cleaner instinct: change what the model is rewarded for, then measure it with metrics that don’t lie. SMOTE isn’t gone. It’s just no longer the answer you reach for first.

Stay ahead, Dan.