Related Neural Networks Links
Learn Evaluation Metrics Neural Networks Tutorial, validate concepts with Evaluation Metrics Neural Networks MCQ Questions, and prepare interviews through Evaluation Metrics Neural Networks Interview Questions and Answers.
Neural Networks
15 Essential Q&A
Interview Prep
Evaluation Metrics — 15 Interview Questions
Accuracy pitfalls, precision/recall trade-offs, ROC vs PR curves, and choosing metrics for imbalanced or multi-class problems.
Colored left borders per card; green / amber / red difficulty chips.
Accuracy
F1
AUC
Confusion
1 Define accuracy.Easy
Answer: (TP+TN) / total—fraction of correct predictions. Misleading when classes are imbalanced (always predicting majority class).
2 Precision vs recall.Easy
Answer: Precision = TP/(TP+FP)—of predicted positives, how many right. Recall = TP/(TP+FN)—of actual positives, how many found. Usually a trade-off via threshold.
Precision = TP/(TP+FP), Recall = TP/(TP+FN)
3 F1 score.Easy
Answer: Harmonic mean of precision and recall: 2PR/(P+R)—penalizes models that are strong on only one; common for imbalanced binary tasks.
4 Confusion matrix.Easy
Answer: Rows/columns for true vs predicted classes; read off TP, FP, TN, FN for binary; extends to multi-class with diagonal = correct.
5 ROC curve and AUC.Medium
Answer: Plot TPR vs FPR as threshold varies. AUC = area—ranking quality; 0.5 random, 1 perfect. Useful when you care about discrimination across thresholds.
6 PR curve vs ROC when imbalanced.Medium
Answer: PR curve (precision vs recall) often more informative with rare positives—ROC can look optimistic because FPR is dominated by negatives.
7 Macro vs micro F1 (multi-class).Hard
Answer: Macro: average F1 per class equally—emphasizes rare classes. Micro: aggregate TP/FP/FN globally then compute P/R/F1—dominated by frequent classes.
8 Top-k accuracy (classification).Easy
Answer: Correct if true label in model’s top k predictions—softer than top-1; used in ImageNet-style benchmarks.
9 Regression: MAE vs RMSE.Medium
Answer: MAE = mean |error|—robust, same units as target. RMSE penalizes large errors more—sensitive to outliers.
10 Log loss (cross-entropy) as metric.Medium
Answer: Penalizes confident wrong probabilities—measures calibration + discrimination; better than accuracy when you need probabilistic quality.
11 Brier score (one line).Hard
Answer: Mean squared error between predicted probability and outcome—for binary, measures calibration and sharpness together.
12 mAP in detection (concept).Hard
Answer: Average precision per class over IoU thresholds / recall levels, then mean across classes—standard object-detection summary metric.
13 Choosing classification threshold.Medium
Answer: Tune on validation to maximize F1, meet minimum recall, or align with business cost FP vs FN—not always 0.5.
14 Why report a naive baseline?Easy
Answer: 95% accuracy means nothing if majority class is 94%—majority classifier sets the floor to beat.
15 Which metric for fraud detection (brief)?Medium
Answer: Often optimize recall at fixed precision or PR-AUC—few positives, cost of FN high; accuracy alone misleading.
Tie metric to business cost of FP vs FN.
Quick review checklist
- Accuracy limits; precision, recall, F1; confusion matrix.
- ROC-AUC vs PR; macro vs micro; threshold choice.
- Regression MAE/RMSE; log loss; know your baseline.