Free DY0-001 practice test: 30 questions with answers

Limiting depth or pruning reduces the variance of a tree by restricting how closely it fits the training data. Increasing depth makes overfitting worse, and linear regression on class labels is not a fix for variance.

Random forests combine bagging with random feature selection, so averaging many decorrelated trees lowers variance. The second option describes the opposite of random forest, and boosting is a different ensemble method.

Gradient boosting fits each new weak learner to the negative gradient of the loss, which for squared error is the residual. Training on labels directly, random features, or cluster assignments does not describe the method.

Silhouette scores measure how well each point fits its own cluster compared with the nearest other cluster, which helps compare values of k. Confusion matrices and early stopping need labels or a supervised training loop.

L1 and L2 penalties both add a term to the loss that grows with coefficient size, with L1 producing sparsity. Adding polynomial terms increases complexity, and removing a validation set does not regularize anything.

Recall measures the share of true cases caught, which matches a setting where false negatives are the costliest error. Precision emphasizes avoiding false alarms, and mean absolute error is a regression metric.

Change data capture reads committed changes from the transaction log and streams them downstream, so only deltas move. Full dumps and manual exports are batch processes that repeat full loads, and UI scraping is fragile and unreliable.

A model registry stores versioned model artifacts, lineage, metrics, and lifecycle stages such as staging and production. A message broker, file format, and dashboard library do not manage model promotion.

Duplicate keys on one side multiply matching rows, which inflates sums and counts. Deduplicating or otherwise enforcing a unique key before the join fixes the fan-out, while cross joins make the problem much worse.

CRISP-DM begins with business understanding, which clarifies objectives and success criteria before any data work. Preparation, modeling, and deployment all depend on that framing.

Shadow deployments and canary releases let teams compare a new model against live behavior before full rollout and roll back quickly. Untracked training and hard-coded notebook thresholds make releases less reproducible and riskier.

A data contract with schema validation checks names, types, and constraints at the boundary, so drift is rejected or flagged immediately. Sampling, compression, and dashboard filters do not enforce the structure of incoming data.

A p-value is the probability of observing a result at least as extreme as the sample result assuming the null hypothesis is true. It does not give the probability that the null or alternative hypothesis is true, and it does not measure effect size.

The Poisson distribution models the number of independent events in a fixed interval at a constant average rate. The exponential distribution models the time between such events, and the binomial needs a fixed number of trials.

The principal components are the eigenvectors of the covariance matrix, ordered by eigenvalue, and the first one points along the direction of greatest variance. Row norms and gradients do not describe variance direction.

Differencing removes a stochastic or linear trend and is the d parameter in ARIMA. Scaling changes the range of values but leaves the trend intact, and one-hot encoding and bootstrapping do not address stationarity.

The posterior probability of disease given a positive test follows from Bayes theorem and depends on prevalence as well as test characteristics. Sensitivity alone ignores false positives in a low-prevalence population.

The gradient points in the direction of steepest increase, so stepping against it reduces the loss locally. Moving along the gradient increases loss, and random steps are not gradient descent.

Features built from post-outcome data and preprocessing fitted before splitting both let information from the validation set or future leak into training. Slightly higher training accuracy and a lower learning rate are normal and do not indicate leakage.

A histogram bins a continuous variable and shows its shape, skew, and tails. Pie charts and stacked areas are for parts of a whole or trends, and network graphs show relationships between entities.

When the probability of missingness depends on the unobserved value itself, the data is missing not at random. Simple mean imputation would bias the estimate in that case, so the analyst should model the missingness or collect more context.

Stakeholders need the decision-relevant range and the assumptions behind it, stated in plain language. Raw coefficients, unqualified claims, and tuning logs do not communicate uncertainty clearly.

Distance-based models are sensitive to feature magnitude, so standardizing or normalizing features prevents income from dominating the distance. Binarizing or removing features loses information and sorting rows does not affect distances.

Using the test set to choose among models leaks information into the selection, so the reported score overstates real performance. Tuning on a validation set or with nested cross-validation keeps the test set as an unbiased final check.

With heavy class imbalance, accuracy can be high even for a useless model, while precision-recall analysis shows how many flagged cases are fraud and how many fraud cases are caught. Mean squared error and R-squared are regression measures.

TF-IDF multiplies term frequency in a document by an inverse document frequency that down-weights terms appearing in many documents. Dense learned vectors are word or sentence embeddings, not TF-IDF.

The reward is scalar feedback received after actions, and the policy defines which action to take in each state. Labeled answers belong to supervised learning, and the learning rate is an optimizer hyperparameter, not a set of states.

Convolutional layers slide learned filters across the image to detect local patterns, which is why they work well for images. Recurrent layers are designed for sequences, and a softmax at the input would have no features to classify.

Collaborative filtering predicts preferences from the behavior of similar users and items. The other options address unrelated problems such as resource allocation or store lifetimes.

Object detection returns the location of each object, usually as a bounding box, together with its class. Image classification gives one label for the whole image, so it cannot localize individual vehicles.