Exam-style
A bank is preparing a training set for a fraud-detection model. It holds 200,000 legitimate transactions and 1,100 fraudulent ones. Separately, a review of a 600-row sample finds that the two annotators who tagged the historical rows disagreed with each other on roughly one label in six, mostly on disputed chargebacks that one treated as fraud and the other as a billing error. What should the team do before training?
Reveal the answer
D. Treat the two findings as separate defects: resample or reweight to address the imbalance, and run an adjudication and relabelling pass to address the disagreement
Two defects sit in this dataset and no single action repairs both. The first is class imbalance: 1,100 positives against 200,000 negatives is a ratio near 1 in 180, so a model that predicts “not fraud” every time scores over 99 per cent accuracy and catches nothing. That is a property of how many rows each class has, it is measurable before training (Amazon SageMaker Clarify reports pre-training bias metrics such as class imbalance and difference in proportions of labels across a facet), and the repairs are the familiar three: resample, by cutting the majority rows or oversampling the minority ones; reweight, so each fraud row counts for more in the loss; or go and collect more positive examples, which is the slower fix and usually the best one. Aiming at balanced datasets is a recognised characteristic of a dataset fit to train on, alongside inclusivity, diversity and curated data sources. The second defect is label quality, and nothing in that list touches it. One label in six disputed means roughly 180 of the 1,100 fraud rows may be filed wrong, and resampling those rows keeps every error it copies; oversampling multiplies them. The repair is adjudication of the disputed rows by a third reviewer, a written annotation guideline that settles what a disputed chargeback counts as, and a relabelling pass over the affected history. Analyzing label quality is named as a tool for detecting bias in its own right, and it is the only one of these that improves the ground truth the model is measured against. That rules out the fifth option, which merges the two, and the first two options, which are model-shaped answers to a data-shaped problem: capacity and regularisation change how well the model fits whatever it is shown, and neither adds a fraud example or corrects a mislabelled one. The third option is the trap worth naming. Deleting age, postcode and account tier is fairness-through-unawareness, and it fails twice. Other columns act as proxies (spend pattern, device, merchant mix, tenure), so the skew survives the deletion, and the attribute just deleted is the one needed to check whether any harm is occurring, which makes the bias unmeasurable rather than absent. Keep those columns, restrict who can read them, and use them for measurement. The check that would have surfaced both defects before anyone opened a training job is subgroup analysis: slice the evaluation results by cohort and read the fraud recall per cohort, rather than reading one aggregate accuracy figure that averages a small group into invisibility. Fairness is a claim about the slices, not about the total.
Q. 1,100 fraud rows against 200,000 legitimate ones, and two annotators who disagreed on one label in six. What gets fixed first?
A. Both, separately. Resampling or reweighting addresses the imbalance; only adjudication and a relabelling pass addresses the labels.
Why? Class imbalance is a count of rows per class, visible before training in SageMaker Clarify’s pre-training bias metrics, and repaired by resampling, reweighting or collecting more positives. Label disagreement is a defect in the ground truth, and resampling carries every wrong label through with it. Deleting the protected attributes is fairness-through-unawareness: proxies keep the skew, and the deletion removes the ability to measure it. Slice the results by cohort and both defects show up before training does.