第 1 / 2 页
非代码面试题
显示 20 / 24 道匹配题目
答题状态：未尝试未正确已正确
ID题目领域难度题型进度权限
2446Hidden Validation Positives Implied by a Leaky Target EncoderA category appears 40 times in train with 18 positives and 10 times in validation. A target encoder is incorrectly fit on train plus validation and outputs 0.56 for that category. How many validation positives did the encoder implicitly use?机器学习简单数值题未尝试免费2447Shift in the Training Mean After Fitting a Scaler on All RowsA feature has training mean 10 over 80 rows and test mean 14 over 20 rows. A scaler is wrongly fit on all 100 rows and uses standard deviation 5. What is the average standardized value of the training block under that leaked fit?机器学习简单数值题未尝试免费2448Held-Out Base Rate Implied by a Full-Sample Class WeightA training set has 100 labels with 30 positives. A class-weighting routine is mistakenly fit on all 125 labels and reports an overall positive rate of 0.36. What is the positive rate in the 25 held-out labels?机器学习中等数值题未尝试面试订阅2449Issuer Demeaning That Quietly Uses Held-Out RowsFor one issuer, the three training rows sum to 12. A pipeline mistakenly demeans by the full-sample issuer mean 3.6 computed from five rows total. What is the sum of the two held-out rows for that issuer?机器学习中等数值题未尝试面试订阅2451Rare Category Survival Caused by Held-Out RowsA categorical preprocessor keeps a level only if it appears at least 5 times. In train alone, level Z appears 4 times. After the preprocessor is wrongly fit on the full sample, level Z appears to have frequency 7 and is kept. How many held-out Z rows caused the leak?机器学习简单数值题未尝试免费2452Future Restatements Merged Into Historical FeaturesA researcher joins fundamentals after they were restated months later, then backtests on the original trade dates. Why is this a split-discipline failure even if no test labels were touched?机器学习中等essay未尝试面试订阅2453Winsor Caps Chosen on the Full PanelA pipeline clips a spread feature at the 1st and 99th percentiles computed on the full panel before the split. Why can this still bias the reported test score even though the clipping rule is unsupervised?机器学习中等essay未尝试面试订阅2454Feature Screening Before the SplitA team ranks 5,000 candidate features by correlation with the target on the full dataset, keeps the top 30, and only then creates train and test. Why is the later split not enough to rescue the experiment?机器学习中等essay未尝试面试订阅2455Repeated Validation Peeking During ResearchA researcher keeps trying new transformations and only retains the ones that improve the same validation score. Why does the validation set stop being a clean model-selection tool?机器学习困难essay未尝试面试订阅2456Row Split Instead of Issuer SplitEach issuer contributes many dated observations. Why can a random row split overstate performance even when the target is defined separately on each date?机器学习简单essay未尝试免费2457PCA Fit Once Before Cross-ValidationA notebook computes PCA on the full feature matrix and then feeds the resulting components into every cross-validation fold. Why is that not a harmless speed optimization?机器学习简单essay未尝试免费2458Choosing Early Stopping by the Test CurveA team trains one model, plots test loss by boosting round, and reports the round with the best test value. Why is the final test score no longer a valid final check?机器学习中等essay未尝试面试订阅2459Using Revised Index Membership in Historical FilteringA backtest filters the universe using current index membership and then evaluates historical predictions on that restricted universe. Why is this also a train/test discipline problem?机器学习困难essay未尝试面试订阅2460Validation Used Until One Model Wins by LuckTwo candidate models are close. A researcher keeps slightly changing seeds and preprocessing until one model wins on the same validation slice. Why should the apparent win be discounted?机器学习困难essay未尝试面试订阅2461Learning Rare-Category Merges From Future FeaturesNo labels are used, but the preprocessing step decides which rare sectors to merge by looking at category frequencies on the full dataset. Why can that still make the evaluation optimistic?机器学习简单essay未尝试免费2462Peer Average Features That Include Held-Out TargetsA feature for each bond is the average realized default rate of bonds from the same issuer-year bucket, computed over the full sample. Why is this worse than ordinary scaling leakage?机器学习中等essay未尝试面试订阅2463Reusing the Test Set After DebuggingA model is evaluated on test, a bug is found, the code is fixed, and the same test set is used again to verify the fix and choose among two corrected versions. Why is that second use no longer a clean test?机器学习中等essay未尝试面试订阅2464No Test Labels Touched Is Not EnoughSomeone argues there was no leakage because the code never accessed test labels. Give the core reason this defense can fail in real ML pipelines.机器学习困难essay未尝试面试订阅2465Why Nested Validation ExistsIf the same validation set is repeatedly used for model family choice, feature engineering, and threshold tuning, why is a second outer holdout or nested procedure conceptually necessary?机器学习困难essay未尝试面试订阅2466What to Audit in a Leakage ReviewYou are auditing a pipeline for leakage. Beyond checking the split line in the final dataframe, what is the highest-value thing to inspect in the code path?机器学习简单essay未尝试免费