DY0-001

A synthetic data set is only useful if it reproduces the statistical properties of the real population it stands in for; if it is not representative, any model trained or validated on it will be biased and will not generalize. Resource use, feature count and row count are practical concerns that can be tuned, since synthetic generation can produce as many rows or features as needed. Representativeness is the core risk.

R-squared (the coefficient of determination) measures the proportion of variance in the dependent variable explained by the regression model, so maximizing it means explaining the most variation. Accuracy and AUC are classification metrics, not regression metrics. A p-value measures statistical significance and should be small, not maximized.

The dependent variable is a real number and the scatter plot shows a strong relationship with a single, normally distributed predictor with few outliers, which fits the assumptions of simple linear regression. Its slope coefficient is directly interpretable as the change in the outcome per unit change in the predictor. Logistic and probit regression model binary outcomes, and exponential regression adds a transformation that makes interpretation less direct without evidence of a nonlinear pattern.

The Akaike information criterion (AIC) compares models fitted by maximum likelihood, trading goodness of fit against the number of parameters, and it does not require the models to be linear or nested, so it suits comparing several nonlinear models. ANOVA-based comparison and the chi-squared test assume specific (often nested or linear) structures or test independence and goodness of fit rather than ranking competing models. MCC (Matthews correlation coefficient) is a single classification metric from a confusion matrix and only applies to classifiers, not to nonlinear models in general.

The depth of a deep neural network is the number of hidden layers between the input and output layers; stacking many hidden layers is what makes a network deep. Convolution and pooling are specific layer types used in convolutional networks, and dropout is a regularization technique, so none of them on its own defines depth.

Scheduling problems assign resources and time slots while respecting constraints such as availability, capacity and precedence, and then optimize an objective such as total time or cost, which is exactly what constrained optimization (for example linear or integer programming) does. A one-armed bandit addresses exploration versus exploitation, a decision tree is a predictive model, and gradient descent is an unconstrained optimization routine for fitting model parameters.

In distributed computing, the workload and data are partitioned across the nodes of a cluster, so adding nodes increases the total memory available and directly relieves a memory constraint. Containerizing or moving to the cloud changes how or where the workload is hosted but does not by itself add memory. Edge deployments typically have fewer resources, making memory constraints worse.

Parquet is a compressed, columnar binary format, so files are small and fast to read, and it stores the schema, including data types, in the file metadata. JSON and XML are verbose text formats that do not enforce data types, and CSV is compact text but carries no schema or data types at all.

K-nearest neighbors is a supervised method that labels a new point by a majority vote of its k closest labeled training samples, so it is used for classification (or regression). K-means is unsupervised: it partitions unlabeled data into k clusters, each represented by a centroid. Option C reverses the two techniques, and neither algorithm is restricted to the data types or data set kinds named in A and B.

A flat tire is an event that happens at a specific moment, while the tire pressure is only measured once a week, so the predictor and the outcome are recorded at different granularities. A pressure drop that leads to a flat can happen entirely between two weekly readings, which makes the measurements too coarse to predict the event; this granularity mismatch is the first thing to resolve. Lagged observations are a modeling technique for time-series features rather than the immediate problem here, and nothing in the scenario points to outliers or a lack of domain expertise.

A greedy algorithm makes the choice that looks best at each step, the locally optimal decision, without revisiting earlier choices; decision-tree induction choosing the best split at each node is the classic example. Updating priors describes Bayesian methods, examining every node is exhaustive search (the opposite of greedy), and applying a theoretical distribution describes parametric modeling.

A REST API exposes the model over standard HTTP methods and URLs, so any department can call it with common tools and libraries and little integration work. SOAP requires XML envelopes and WSDL contracts, and RPC frameworks usually need generated client stubs, both of which add development effort. JSON is a data format that REST APIs often use, not an API style.

A container image packages the application with only the libraries and dependencies it needs and shares the host operating system kernel, so it is lightweight, portable and reusable across environments. A virtual machine packages a full guest operating system, which is far more than the critical dependencies. Thin clients and edge devices are endpoint hardware, not packaging models.

N-grams capture sequences of adjacent words, so they show which terms occur together and let the analyst associate words into concepts (for example, "credit" and "card" forming "credit card"). Named entity recognition only tags names of people, places and organizations, TF-IDF weights how important individual terms are to a document, and part-of-speech tagging labels the grammatical role of each word; none of these builds associations between terms.

Executives need to know what was found and what to do about it, so a presentation to them should give high-level results and final recommendations. Literature reviews, code snippets and detailed explanations of statistical tests are technical detail for a technical audience or an appendix. Security keys and login information must never appear in any presentation.

A scatter plot shows one variable against another, so it reveals relationships such as linearity, curvature and clusters between them. Violin and box-and-whisker plots summarize the distribution of a single variable, and a Q-Q plot compares a variable's distribution with a theoretical one such as the normal distribution rather than showing the relationship between two variables.

A scheduled bus is aimed at a fixed time each hour, so its actual arrival times cluster symmetrically around the scheduled time with small random deviations early or late, which a normal distribution models well. Exponential and Poisson describe random, unscheduled events: the waiting time between independent arrivals and the count of arrivals in an interval. Binomial counts successes in a fixed number of yes/no trials, not a continuous time.

The model already meets the agreed requirements, so whether to trade the deadline for more accuracy is a business decision that belongs to the key stakeholder. Consulting them first sets expectations and decides if extra work is worth it. Collecting more data, testing more specifications or asking for funding all extend scope or cost without that agreement.

Ridge regression minimizes the sum of squared residuals (e²), the ordinary least squares loss, plus an L2 penalty on the size of the coefficients. The absolute value |e| is the loss of least absolute deviations regression, and the L1 penalty belongs to lasso, not ridge. Minimizing the raw residual e is meaningless because positive and negative errors cancel, and zero residuals is not an achievable objective.

Binning groups a continuous variable such as age into ranges, so sparse observations with gaps are aggregated into larger, more stable groups. Imputing fills in missing values rather than aggregating observations, label encoding turns categories into numbers, and linearization transforms a relationship to make it linear.