CompTIA DataAI (DataX) DY0-001 Practice Exam

Correct: B — AUC. ROC/AUC summarises a classifier’s separability across all thresholds in one number, so it is threshold-independent by construction — exactly what the requirement asks. It is a named statistical-methods objective.

Why not the others: Accuracy at 0.5 (A) is tied to a single threshold, the opposite of the requirement, and is misleading on imbalanced data. Mean squared error (C) is a regression metric, not a classifier-separation measure. A t-test p-value (D) tests a hypothesis about means, not classification quality.

Correct: C — overfitting; regularize or simplify. A large train-to-validation gap is the signature of overfitting — the model has memorised the training data and generalises poorly. Regularization, reducing complexity, or gathering more data are the standard first remedies.

Why not the others: Underfitting (A) shows as poor performance on both sets, and adding complexity would worsen this case. Data leakage (B) usually inflates validation scores too, and it always needs fixing. Class imbalance (D) is a different problem, and one would oversample the minority, not majority, class.

Correct: A — a single decision tree. A single tree yields a directly readable path of if-then splits for any prediction, which is the transparency the auditors require. The scenario explicitly accepts lower accuracy for that interpretability.

Why not the others: A deep network (B) is a black box without extra explainability tooling. An RBF-kernel SVM (C) maps data into a high-dimensional space that resists simple explanation. A boosted ensemble (D) is accurate but combines hundreds of trees, so no single readable path exists.

Correct: C — PCA. Principal component analysis projects correlated features onto a smaller set of orthogonal components that retain most of the variance, which is exactly the dimensionality-reduction goal. It is a named dimensionality-reduction technique in the objectives.

Why not the others: One-hot encoding (A) expands categorical variables into more columns — it increases dimensionality. Min-max scaling (B) rescales values but keeps all 40 features. SMOTE (D) addresses class imbalance by synthesising minority samples, not reducing features.

Correct: D — it is data leakage. Fitting scaling parameters on all the data lets statistics from the validation folds leak into preprocessing, so the cross-validation score is optimistically biased. Scaling must be fit on the training fold only, inside the CV loop.

Why not the others: “Scaling has no effect” (A) is false for distance- and gradient-based models. Underfitting (B) is unrelated — the issue is an inflated, not degraded, estimate. “Only tree-based models” (C) is backwards: trees are relatively scale-insensitive, so the leakage concern applies more to other model families.

Correct: A — continuous monitoring for drift. Ongoing performance and drift monitoring is the MLOps practice built to detect gradual degradation as input distributions change, triggering retraining before the impact grows. It is an explicit operations-domain objective.

Why not the others: A one-time acceptance test (B) validates the model at launch but sees nothing afterwards. Enlarging the training set once (C) helps initial fit but does not detect later drift. Code review (D) checks the script, not the live model’s behaviour over time.

Correct: B — version control, data lineage, and containers. Reproducibility requires pinning all three moving parts: the exact code (version control), the exact data (lineage/versioning), and the exact runtime (containerization). The data-science life-cycle and deployment objectives name each of these.

Why not the others: Saving only the model file (A) loses the code and data needed to regenerate it. Emailing a notebook (C) captures neither data version nor environment. Re-running current production (D) uses today’s code and data, not the six-month-old state.

Correct: D — TF-IDF. TF-IDF scales a term’s frequency in a document by how rare it is across the corpus, so ubiquitous words are down-weighted and distinctive words up-weighted — exactly the described behaviour. It is a named NLP technique in the objectives.

Why not the others: One-hot encoding (A) marks presence without any corpus-wide weighting. OCR (B) extracts text from images and is unrelated to term weighting. Batch normalization (C) is a deep-learning training technique, not an NLP weighting scheme.

Correct: B — data augmentation. Augmentation synthesises new training views — flips, rotations, crops, colour shifts — from existing images, expanding effective dataset size and improving robustness without new data collection. It is a named computer-vision objective.

Why not the others: Lowering resolution (A) discards information and typically hurts accuracy. Changing the loss to MAE (C) is a regression choice unrelated to limited image data. Removing convolutional layers (D) strips out the very structure that makes vision models effective.

Correct: C — recall/precision on the minority class; the issue is class imbalance. With 0.5% fraud, a “always legitimate” model scores high accuracy while catching zero fraud — its recall on the fraud class is 0. Precision and recall on the minority class expose this, and remedies include resampling or class weighting.

Why not the others: Accuracy (A) is exactly the misleading metric the scenario warns about on imbalanced data. MSE (B) and R-squared (D) are regression metrics and do not apply to this classification problem, and neither names the imbalance that is the real issue.