DY0-001 Sample Questions & Answers

t-SNE is a non-linear dimensionality reduction technique specifically designed for visualizing high-dimensional data in low-dimensional space (typically 2D or 3D). It excels at revealing the underlying structure of data, such as clusters, by preserving the local similarities between data points. PCA, in contrast, is a linear technique focused on maximizing variance and may not effectively separate clusters that are not linearly separable.

A right-skewed (or positively skewed) distribution has a long tail extending to the right. This is caused by a smaller number of high-value outliers pulling the mean to the right. In such a distribution, the typical relationship is Mean > Median > Mode. The median is less affected by outliers than the mean, so it remains closer to the bulk of the data.

Model drift, also known as concept drift, occurs when the statistical properties of the target variable, which the model is trying to predict, change over time in unforeseen ways. This causes the model, which was trained on historical data, to become less accurate as time passes. The gradual decrease in performance is a classic symptom of model drift, often caused by changes in fraudulent behavior patterns.

This represents a standard and effective pipeline for text classification. 1) Tokenization breaks the text into individual words (tokens). 2) Stop word removal eliminates common words ('the', 'a', 'is') that add little semantic value. 3) TF-IDF (Term Frequency-Inverse Document Frequency) vectorization converts the cleaned text into numerical vectors that represent the importance of each word in the context of the entire corpus, which is an ideal input for most machine learning classifiers.

This is the correct definition of the R-squared (coefficient of determination) value. It is a statistical measure that provides a 'goodness of fit' for the model, indicating how much of the variability in the outcome data can be explained by the model's inputs.

Logistic Regression is a linear model that is highly interpretable. The coefficients of the model directly indicate the influence of each feature on the predicted outcome's log-odds, making it easy for clinicians to understand which factors are most significant.

Decision Trees are inherently interpretable, often called 'white-box' models. The tree structure provides a clear set of if-then rules that can be easily visualized and followed to understand how a specific prediction was made, which is ideal for clinical review.

This statement is the correct definition of the Central Limit Theorem (CLT). It is a fundamental concept in statistics that allows for making inferences about a population mean based on a sample mean, even if the population itself is not normally distributed, provided the sample size is large enough (commonly n > 30).

Target Encoding (or Mean Encoding) is highly effective for high-cardinality categorical features, especially with tree-based models. It replaces each category with the mean of the target variable for that category. This captures information about the target variable directly in the feature, avoiding the massive dimensionality increase of One-Hot Encoding and the arbitrary ordering problem of Label Encoding. Proper implementation requires careful handling (e.g., using cross-validation) to prevent data leakage.

SARIMA is an extension of the ARIMA model specifically designed for time series data that contains a seasonal component. It includes additional parameters to model the seasonality alongside the trend and autoregressive components. Given the strong seasonality and trend in the sales data, SARIMA is the most appropriate and powerful choice among the options.

Elastic Net is the best choice here because it addresses all key requirements. As a regression model, it predicts a continuous value. Its key advantage is the combination of L1 (Lasso) and L2 (Ridge) regularization. L2 regularization helps prevent overfitting and handles multicollinearity well. L1 regularization can perform automatic feature selection by shrinking some feature coefficients to zero, which directly addresses the need for understanding feature importance. This combination makes it a robust and interpretable choice for this complex business problem.