Feature hashing<\/span><\/p>\nAnswer: 1) Stacking<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Stacking is an ensemble learning technique that combines predictions from multiple base models by using another model, called a meta-model, to produce the final prediction. The base models can use different algorithms or feature representations, allowing the ensemble to capture different patterns in the data. The meta-model learns how to combine their outputs effectively. This differs from simple voting, where predictions are combined directly without training a separate meta-model. Stacking can improve predictive performance when the base models make complementary errors. Proper validation is important so that the meta-model does not learn from predictions generated using data that caused leakage.<\/span><\/p>\nQuestion 342. In a classification ensemble, what does hard voting generally mean?<\/b><\/h3>\n
1)<\/b> Averaging predicted probabilities<\/span>
\n<\/span>2)<\/b> Selecting the class receiving the most model votes<\/span>
\n<\/span>3)<\/b> Training a neural network on ensemble outputs<\/span>
\n<\/span>4)<\/b> Selecting the prediction with the highest confidence score from one model<\/span><\/p>\nAnswer: 2) Selecting the class receiving the most model votes<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Hard voting combines predictions from multiple classification models by selecting the class that receives the largest number of votes. For example, if five classifiers predict three classes and three classifiers select Class A, while the remaining two select Class B, the ensemble selects Class A. Hard voting generally uses the predicted class labels rather than the probability values. This makes it different from soft voting, which typically combines predicted probabilities. Hard voting can be useful when several reasonably strong and diverse models are available. The effectiveness of the approach depends on the quality, diversity, and independence of the participating models.<\/span><\/p>\nQuestion 343. What does the bias-variance tradeoff describe in machine learning?<\/b><\/h3>\n
1)<\/b> The relationship between training data size and storage requirements<\/span>
\n<\/span>2)<\/b> The difference between classification and regression<\/span>
\n<\/span>3)<\/b> The balance between model simplicity and sensitivity to training data<\/span>
\n<\/span>4)<\/b> The relationship between CPU usage and prediction latency<\/span><\/p>\nAnswer: 3) The balance between model simplicity and sensitivity to training data<\/b><\/p>\n
Explanation:<\/b><\/p>\n
The bias-variance tradeoff describes the balance between two important sources of prediction error. A model with high bias is generally too simple and may fail to capture meaningful patterns in the data. A model with high variance can respond too strongly to particular training examples and may perform poorly on unseen data. Increasing model complexity can reduce bias but may increase variance. Conversely, simplifying a model may reduce variance while increasing bias. The objective is not to eliminate one completely, but to achieve a useful balance that provides reliable generalization. Validation data is commonly used to assess how well this balance is achieved.<\/span><\/p>\nQuestion 344. What is irreducible error in a machine learning problem?<\/b><\/h3>\n
1)<\/b> Error caused by incorrect model deployment<\/span>
\n<\/span>2)<\/b> Error caused by insufficient hyperparameter tuning<\/span>
\n<\/span>3)<\/b> Error caused by an overly complex model<\/span>
\n<\/span>4)<\/b> Error resulting from inherent noise or uncertainty in the data<\/span><\/p>\nAnswer: 4) Error resulting from inherent noise or uncertainty in the data<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Irreducible error is the portion of prediction error that cannot be eliminated simply by selecting a different machine learning model. It can result from measurement noise, missing information, random variation, or uncertainty that is inherently present in the problem. For example, two observations with identical measured features may still have different outcomes because important variables are unavailable or because the outcome contains natural randomness. A more complex model cannot reliably remove this type of uncertainty. Understanding irreducible error is useful because it establishes a practical limit on how accurately a prediction system can perform, even when the model and training process are well designed.<\/span><\/p>\nQuestion 345. What is the main difference between a model parameter and a hyperparameter?<\/b><\/h3>\n
1)<\/b> Parameters are learned from training data, while hyperparameters are set before or during the training configuration<\/span>
\n<\/span>2)<\/b> Parameters are stored only in S3, while hyperparameters are stored only in CloudWatch<\/span>
\n<\/span>3)<\/b> Parameters are used only for classification, while hyperparameters are used only for regression<\/span>
\n<\/span>4)<\/b> Parameters cannot change during training, while hyperparameters are always learned automatically<\/span><\/p>\nAnswer: 1) Parameters are learned from training data, while hyperparameters are set before or during the training configuration<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Model parameters are values learned by an algorithm from training data. Examples include weights in a neural network and coefficients in many regression models. Hyperparameters are configuration values that control the learning process or model structure and are normally specified before training or selected through a tuning process. Examples include learning rate, batch size, tree depth, and regularization strength. During training, the algorithm updates parameters based on the data, but hyperparameters are not normally learned in the same direct manner. Understanding this distinction is important because hyperparameter optimization focuses on finding suitable configuration values that help the model achieve better validation performance.<\/span><\/p>\nQuestion 346. What is the primary characteristic of grid search for hyperparameter optimization?<\/b><\/h3>\n
1)<\/b> It selects hyperparameters using a neural network<\/span>
\n<\/span>2)<\/b> It evaluates predefined combinations from the specified hyperparameter grid<\/span>
\n<\/span>3)<\/b> It always evaluates only one configuration<\/span>
\n<\/span>4)<\/b> It chooses configurations without using validation results<\/span><\/p>\nAnswer: 2) It evaluates predefined combinations from the specified hyperparameter grid<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Grid search evaluates combinations of hyperparameter values defined by the user. For example, a practitioner might specify several learning rates and several batch sizes, creating a set of combinations to evaluate. Each configuration can be trained and measured using a validation metric, and the results can then be compared. Grid search is straightforward and systematic, but the number of combinations can grow rapidly as more hyperparameters and values are included. This can increase computational cost. It is most practical when the search space is relatively small or carefully restricted. The selected configuration should generally be evaluated on an untouched test dataset only after tuning is complete.<\/span><\/p>\nQuestion 347. Why can random search be more efficient than grid search for some hyperparameter optimization problems?<\/b><\/h3>\n
1)<\/b> It guarantees the globally optimal configuration<\/span>
\n<\/span>2)<\/b> It does not require any validation data<\/span>
\n<\/span>3)<\/b> It can explore more varied combinations without evaluating every grid point<\/span>
\n<\/span>4)<\/b> It automatically removes irrelevant training features<\/span><\/p>\nAnswer: 3) It can explore more varied combinations without evaluating every grid point<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Random search samples hyperparameter configurations from specified distributions or ranges instead of systematically evaluating every combination in a predefined grid. This can be more efficient when only a few hyperparameters have a strong influence on model performance. Grid search may spend many evaluations on combinations that contribute little useful information. Random search can distribute trials across a broader search space and may discover useful regions with fewer evaluations. It does not guarantee the optimal configuration, however. The number of trials, parameter distributions, validation strategy, and computational budget all affect its effectiveness. Random search is particularly useful when the hyperparameter space contains many dimensions.<\/span><\/p>\nQuestion 348. What is a key characteristic of Bayesian optimization for hyperparameter tuning?<\/b><\/h3>\n
1)<\/b> It evaluates every possible hyperparameter combination<\/span>
\n<\/span>2)<\/b> It randomly deletes poorly performing training examples<\/span>
\n<\/span>3)<\/b> It requires all models to use identical architectures<\/span>
\n<\/span>4)<\/b> It uses previous evaluation results to select promising configurations for subsequent trials<\/span><\/p>\nAnswer: 4) It uses previous evaluation results to select promising configurations for subsequent trials<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Bayesian optimization is a sequential strategy for finding effective hyperparameter configurations. Instead of treating every trial independently, it uses information from previous evaluations to build a model or surrogate of the relationship between hyperparameters and objective performance. The optimization process then uses this information to select promising configurations for future trials. This can reduce the number of expensive training jobs required compared with exhaustive approaches. Bayesian optimization is particularly useful when model training is computationally expensive. The method still depends on an appropriate objective metric, search space, and validation process, and it does not guarantee that every trial will improve upon the previous best result.<\/span><\/p>\nQuestion 349. What does a learning curve typically show?<\/b><\/h3>\n
1)<\/b> Model performance as the amount of training data changes<\/span>
\n<\/span>2)<\/b> The AWS cost of each training instance<\/span>
\n<\/span>3)<\/b> The number of features stored in a feature repository<\/span>
\n<\/span>4)<\/b> The network bandwidth used by an endpoint<\/span><\/p>\nAnswer: 1) Model performance as the amount of training data changes<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A learning curve typically illustrates model performance as the amount of training data increases. It commonly compares training and validation performance across progressively larger subsets of the available training dataset. This visualization can provide insight into whether additional data may improve generalization. For example, if validation performance continues improving as more data is added, additional training examples may be useful. If both training and validation performance remain poor, the model may require changes beyond simply adding data. Learning curves therefore help practitioners understand how model behavior changes with dataset size and can support decisions about data collection, model complexity, and training strategy.<\/span><\/p>\nQuestion 350. What is covariate shift in a machine learning system?<\/b><\/h3>\n
1)<\/b> The distribution of target labels changes while the input distribution remains fixed<\/span>
\n<\/span>2)<\/b> The distribution of input features changes while the relationship between inputs and targets remains approximately unchanged<\/span>
\n<\/span>3)<\/b> The model architecture changes during every prediction<\/span>
\n<\/span>4)<\/b> The training dataset is encrypted using a different key<\/span><\/p>\nAnswer: 2) The distribution of input features changes while the relationship between inputs and targets remains approximately unchanged<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Covariate shift occurs when the distribution of input features changes between training and deployment while the conditional relationship between the inputs and target remains approximately stable. In notation, the distribution of X changes while P(Y|X) remains similar. For example, a model trained using customer behavior from one period may encounter a different customer population later, causing feature distributions to change. This can affect model performance even when the underlying relationship between features and outcomes remains stable. Monitoring feature distributions between training and production data can help identify this situation and determine whether retraining or other corrective actions may be necessary.<\/span><\/p>\nQuestion 351. What does label shift refer to in machine learning?<\/b><\/h3>\n
1)<\/b> A change in the names assigned to model features<\/span>
\n<\/span>2)<\/b> A change in the model’s hyperparameter values<\/span>
\n<\/span>3)<\/b> A change in the distribution of target classes while the class-conditional feature distributions remain approximately stable<\/span>
\n<\/span>4)<\/b> A change in the order of records in a dataset<\/span><\/p>\nAnswer: 3) A change in the distribution of target classes while the class-conditional feature distributions remain approximately stable<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Label shift describes a situation in which the distribution of target labels changes between training and deployment, while the distribution of features conditioned on each class remains approximately stable. In other words, P(Y) changes while P(X|Y) is assumed to remain similar. For example, a classification system may have been trained when positive cases represented a relatively small percentage of observations, but the production environment may later contain a substantially different proportion. Recognizing label shift is useful because a model’s predicted probabilities and overall performance can be affected even if the relationship between features and each class remains relatively consistent.<\/span><\/p>\nQuestion 352. How is permutation feature importance generally calculated?<\/b><\/h3>\n
1)<\/b> By deleting the model and retraining it without each feature<\/span>
\n<\/span>2)<\/b> By counting the number of missing values in each feature<\/span>
\n<\/span>3)<\/b> By measuring the storage size of each feature<\/span>
\n<\/span>4)<\/b> By randomly shuffling a feature and measuring the resulting performance decrease**<\/span><\/p>\nAnswer: 4) By randomly shuffling a feature and measuring the resulting performance decrease<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Permutation feature importance measures how much a model depends on a particular feature by disrupting that feature’s values and observing the resulting change in model performance. Typically, the values of one feature are randomly shuffled while the other features remain unchanged. If model performance decreases substantially, the feature likely provides useful predictive information to the model. If performance changes very little, the feature may contribute less to the model’s predictions. The method can be applied to many model types because it evaluates the model’s behavior rather than relying on a specific internal model structure. Correlated features can complicate interpretation.<\/span><\/p>\nQuestion 353. What do SHAP values primarily explain in a machine learning model?<\/b><\/h3>\n
1)<\/b> The contribution of individual features to a specific prediction<\/span>
\n<\/span>2)<\/b> The amount of S3 storage used by the model<\/span>
\n<\/span>3)<\/b> The number of training instances required by an algorithm<\/span>
\n<\/span>4)<\/b> The network latency of an inference endpoint<\/span><\/p>\nAnswer: 1) The contribution of individual features to a specific prediction<\/b><\/p>\n
Explanation:<\/b><\/p>\n
SHAP, or SHapley Additive exPlanations, provides a way to describe how individual features contribute to a model prediction. SHAP values are based on concepts from cooperative game theory and estimate how each feature contributes to moving a prediction away from a reference or baseline value. Positive and negative contributions can indicate how features influence a prediction in different directions. SHAP explanations can be useful for understanding complex models and investigating individual predictions. However, the interpretation depends on the selected background data, model, and explanation method. SHAP values describe model behavior and should not automatically be interpreted as proof of causal relationships.<\/span><\/p>\nQuestion 354. What is the primary purpose of a partial dependence plot?<\/b><\/h3>\n
1)<\/b> To display the raw training records<\/span>
\n<\/span>2)<\/b> To show the average predicted response as a feature varies<\/span>
\n<\/span>3)<\/b> To measure AWS storage consumption<\/span>
\n<\/span>4)<\/b> To replace the model’s validation dataset<\/span><\/p>\nAnswer: 2) To show the average predicted response as a feature varies<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A partial dependence plot, or PDP, helps illustrate the relationship between one or more features and a model’s predicted outcome. For a selected feature, the method varies that feature across a range of values while averaging the model’s predictions over the observed values of other features. The resulting plot shows the model’s average response to changes in the selected feature. PDPs can help interpret complex models and identify nonlinear relationships. However, strong correlations between features can make interpretation more difficult. A partial dependence plot describes how the model behaves; it does not establish that changing a feature would necessarily cause the real-world outcome to change.<\/span><\/p>\nQuestion 355. What does an Individual Conditional Expectation (ICE) plot provide?<\/b><\/h3>\n
1)<\/b> The average prediction for the entire dataset only<\/span>
\n<\/span>2)<\/b> The distribution of model training costs<\/span>
\n<\/span>3)<\/b> Prediction-response curves for individual observations as a feature changes<\/span>
\n<\/span>4)<\/b> A list of the model’s learned parameter values<\/span><\/p>\nAnswer: 3) Prediction-response curves for individual observations as a feature changes<\/b><\/p>\n
Explanation:<\/b><\/p>\n
An Individual Conditional Expectation, or ICE, plot shows how a model’s prediction changes for individual observations as the value of a selected feature is varied. Instead of averaging all observations into one curve, ICE produces a separate prediction curve for each observation. This can reveal differences that an overall partial dependence plot might hide. For example, different groups of customers may respond differently to changes in the same feature according to the model. ICE plots are therefore useful for examining heterogeneous model behavior. As with other model interpretation techniques, correlations, extrapolation, and model-specific assumptions should be considered when interpreting the results.<\/span><\/p>\nQuestion 356. What is the purpose of a champion-challenger approach in machine learning?<\/b><\/h3>\n
1)<\/b> To permanently delete the existing production model<\/span>
\n<\/span>2)<\/b> To train every model using identical data and parameters<\/span>
\n<\/span>3)<\/b> To replace validation with manual inspection<\/span>
\n<\/span>4)<\/b> To compare a current production model with a candidate model using defined evaluation criteria**<\/span><\/p>\nAnswer: 4) To compare a current production model with a candidate model using defined evaluation criteria<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A champion-challenger approach maintains a current production model, called the champion, while evaluating one or more candidate models, called challengers. The challenger can be tested against the champion using predefined metrics, controlled traffic, historical data, or other evaluation methods. The purpose is to establish an evidence-based process for determining whether a candidate model is suitable for replacing the current model. This approach can support continuous model improvement while maintaining a known production baseline. The evaluation should use appropriate datasets and metrics, and deployment changes should follow established testing, approval, monitoring, and rollback procedures.<\/span><\/p>\nQuestion 357. Why is diversity among ensemble models often useful?<\/b><\/h3>\n
1)<\/b> Different models can make different errors, allowing the ensemble to combine complementary predictions<\/span>
\n<\/span>2)<\/b> Diverse models always require less training data<\/span>
\n<\/span>3)<\/b> Diversity guarantees perfect prediction accuracy<\/span>
\n<\/span>4)<\/b> Diverse models eliminate the need for validation<\/span><\/p>\nAnswer: 1) Different models can make different errors, allowing the ensemble to combine complementary predictions<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Model diversity can improve the usefulness of an ensemble because different models may capture different patterns or make different prediction errors. When the errors made by individual models are not highly correlated, combining their predictions can produce a more stable result. For example, two models may have similar overall accuracy but fail on different observations. An ensemble can potentially benefit from their complementary strengths. Diversity alone does not guarantee better performance, however. The individual models must also provide useful predictive information. Ensemble design should therefore consider both model quality and the degree to which their predictions or errors differ.<\/span><\/p>\nQuestion 358. What is weighted voting in an ensemble classification system?<\/b><\/h3>\n
1)<\/b> Giving every model exactly the same influence<\/span>
\n<\/span>2)<\/b> Assigning different importance weights to model predictions when combining them<\/span>
\n<\/span>3)<\/b> Selecting a single model before training begins<\/span>
\n<\/span>4)<\/b> Removing all models with different architectures<\/span><\/p>\n