I85PREP
⌘ K
EXIN BCS ML Award study materials
EXIN BCS ML Award study resources

EXIN BCS ML Award practice questions

One question at a time. Build your EXIN BCS ML Award understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Introduction to Machine Learning

Which statement best describes supervised learning?

  1. The model learns by trial and error using rewards and penalties from an environment.
  2. The model reduces the number of input features before any learning takes place.
  3. The model groups unlabelled data points by similarity without any target variable.
  4. The model learns a mapping from labelled input-output pairs to predict outputs for new inputs. Correct answer.

Explanation

Supervised learning uses a training set in which each input is paired with a known target output, and the algorithm learns a mapping that generalises to unseen inputs. Classification and regression are both supervised tasks. The presence of labelled targets is what distinguishes it from unsupervised and reinforcement approaches.

Question 2Introduction to Machine Learning

Predicting tomorrow's temperature in degrees Celsius is an example of which type of machine learning task?

  1. Clustering
  2. Dimensionality reduction
  3. Classification
  4. Regression Correct answer.

Explanation

Regression tasks predict a continuous numeric value, and temperature in degrees is continuous. Classification predicts a discrete class label, clustering discovers groups in unlabelled data, and dimensionality reduction compresses features. Because the output is a quantity on a continuous scale, regression is the correct framing.

Question 3Introduction to Machine Learning

In machine learning terminology, what is a feature?

  1. A measurable input variable used to describe an example Correct answer.
  2. The number of examples in the training dataset
  3. The error between a prediction and the true label
  4. The target value the model is trained to predict

Explanation

A feature is an individual measurable property or input variable that describes each example, such as age or pixel intensity. The target is the label being predicted, and the error measures prediction quality. Features form the input representation the model learns from.

Question 4Introduction to Machine Learning

What does the term 'generalisation' refer to in machine learning?

  1. The process of converting raw data into numeric features
  2. The speed at which a model processes training examples
  3. The ability of a model to perform well on data it was not trained on Correct answer.
  4. The total number of parameters inside a neural network

Explanation

Generalisation is the capacity of a trained model to make accurate predictions on new, unseen data drawn from the same distribution. It is the central goal of learning, since performance on training data alone does not guarantee usefulness. Overfitting is the failure to generalise.

Question 5Introduction to Machine Learning

A model performs very well on its training data but poorly on new data. Which phenomenon does this describe?

  1. Regularisation
  2. Overfitting Correct answer.
  3. Underfitting
  4. Data leakage

Explanation

Overfitting occurs when a model captures noise and idiosyncrasies of the training set, so its training performance is high while performance on unseen data is poor. It typically arises when the model is too complex relative to the amount of data. Underfitting is the opposite pattern.

Question 6Introduction to Machine Learning

Which of the following is an example of unsupervised learning?

  1. Predicting whether a loan applicant will default
  2. Recognising handwritten digits from labelled images
  3. Customer segmentation using clustering Correct answer.
  4. Estimating house prices from property attributes

Explanation

Clustering customers into segments uses unlabelled data and seeks structure without a target variable, which is unsupervised learning. The other options all rely on labelled examples to predict a known output. The absence of labels is the defining characteristic here.

Question 7Introduction to Machine Learning

Why is a separate test set used when evaluating a machine learning model?

  1. To store the model parameters after training
  2. To speed up the training of the algorithm
  3. To provide an unbiased estimate of performance on unseen data Correct answer.
  4. To reduce the number of features used during training

Explanation

A test set is held out from training so that evaluation reflects how the model behaves on data it has never seen. This gives a more honest estimate of generalisation than training or validation performance. Reusing the test set for tuning would bias that estimate.

Question 8Introduction to Machine Learning

Which statement correctly distinguishes artificial intelligence, machine learning and deep learning?

  1. Artificial intelligence is a subset of machine learning, which is a subset of deep learning
  2. Deep learning is a subset of machine learning, which is a subset of artificial intelligence Correct answer.
  3. They are three names for exactly the same technique
  4. Machine learning is a subset of deep learning, which is a subset of artificial intelligence

Explanation

Artificial intelligence is the broad field of building systems that perform tasks associated with intelligence. Machine learning is a subfield in which systems learn from data, and deep learning is a further subfield of machine learning based on multi-layer neural networks. The nesting is therefore AI containing ML containing DL.

Question 9Introduction to Machine Learning

A team wants to predict whether an email is spam or not spam. Which task type and output does this represent?

  1. Clustering, because emails are grouped by similarity
  2. Reinforcement learning, because the filter improves over time
  3. Regression, because the output is a numeric probability
  4. Binary classification, because there are two discrete classes Correct answer.

Explanation

Spam detection assigns each email to one of two discrete categories, spam or not spam, which is binary classification. Even if the model outputs a probability, the decision is between two classes. Regression would apply to a continuous target.

Question 10Introduction to Machine Learning

What is the main purpose of splitting data into training and validation sets?

  1. To double the amount of data available for fitting the model
  2. To tune model choices and hyperparameters without touching the final test data Correct answer.
  3. To remove outliers from the dataset automatically
  4. To ensure the model cannot be deployed until validation is complete

Explanation

The validation set is used to compare models and tune hyperparameters, guiding development decisions without contaminating the test set. This keeps the test set as an untouched estimate of final performance. Outlier removal and deployment are separate concerns.

Question 11Introduction to Machine Learning

A dataset contains 990 healthy patients and 10 diseased patients. A model that always predicts 'healthy' reports 99% accuracy. What does this illustrate?

  1. That the model has overfitted the training data
  2. That accuracy is a reliable metric for imbalanced data
  3. That regression should be used instead of classification
  4. That accuracy can be misleading when classes are imbalanced Correct answer.

Explanation

When one class heavily dominates, a trivial majority-class predictor can achieve high accuracy while completely failing to detect the minority class. This shows why accuracy alone is inadequate for imbalanced problems and why metrics such as recall or precision are needed. The high score reflects class imbalance, not genuine skill.

Question 12Introduction to Machine Learning

Which technique is most directly aimed at reducing overfitting in a decision tree?

  1. Increasing the maximum depth of the tree
  2. Adding more features to the dataset
  3. Removing the validation set
  4. Limiting the maximum depth of the tree Correct answer.

Explanation

Limiting tree depth is a form of pre-pruning that constrains model complexity, which reduces the tendency to memorise noise. Deeper trees fit training data more closely and overfit more easily. Constraining complexity is a standard remedy for overfitting.

Question 13Introduction to Machine Learning

What is the role of a loss function during model training?

  1. To quantify how far predictions are from the true targets so parameters can be adjusted Correct answer.
  2. To select which features to discard before training
  3. To store the training data in a compressed form
  4. To define the number of layers in a neural network

Explanation

A loss function measures the discrepancy between predicted and actual values, producing a signal that optimisation algorithms use to update model parameters. Minimising this loss is the objective of training. It does not itself select features or define architecture.

Question 14Introduction to Machine Learning

Which description best matches reinforcement learning?

  1. A model learns from a fixed set of labelled examples provided in advance
  2. A model discovers clusters in unlabelled data by measuring distances
  3. A model compresses high-dimensional data into fewer dimensions
  4. An agent learns a policy by interacting with an environment and receiving rewards Correct answer.

Explanation

Reinforcement learning involves an agent that takes actions in an environment and learns from the rewards or penalties it receives, aiming to maximise cumulative reward. Unlike supervised learning, no fixed labelled dataset is provided. The feedback is evaluative rather than a correct answer per example.

Question 15Introduction to Machine Learning

A model is trained to predict customer churn and achieves 80% recall on churners but only 40% precision. What does this mean in practice?

  1. It catches most churners but many of its churn predictions are false alarms Correct answer.
  2. It predicts churn at random with no useful signal
  3. It catches few churners but almost every prediction it makes is correct
  4. It has perfect performance on the churn class

Explanation

Recall of 80% means the model identifies most actual churners, while precision of 40% means only two in five predicted churners truly churn. The model therefore misses few churners but generates many false positives. This trade-off matters when retention actions are costly.

Question 16Introduction to Machine Learning

What is feature engineering?

  1. Measuring the computational cost of training a model
  2. Deploying a trained model into a production environment
  3. Using domain knowledge to create or transform input variables that help the model learn Correct answer.
  4. Automatically selecting the learning algorithm with the best accuracy

Explanation

Feature engineering is the use of domain knowledge to construct, combine or transform input variables so that patterns become easier for the model to learn. Examples include deriving ratios or encoding dates. It is distinct from algorithm selection and deployment.

Question 17Introduction to Machine Learning

Which situation most clearly indicates that a model is underfitting?

  1. Test error is lower than training error by a large margin
  2. Training error is high and test error is also high Correct answer.
  3. Training error is very low and test error is very high
  4. Training error equals test error at a low value

Explanation

Underfitting occurs when a model is too simple to capture the underlying pattern, producing poor performance on both training and test data. Overfitting, by contrast, shows low training error with high test error. High error on both sets signals insufficient model capacity or poor features.

Question 18Introduction to Machine Learning

What is the purpose of cross-validation?

  1. To obtain a more reliable performance estimate by training and testing on multiple data splits Correct answer.
  2. To reduce the number of classes in a classification problem
  3. To guarantee that the model will not overfit
  4. To increase the size of the dataset by generating synthetic examples

Explanation

Cross-validation partitions the data into several folds and repeatedly trains and evaluates, averaging results across splits. This yields a more stable estimate of performance than a single split, especially with limited data. It reduces variance in the estimate but does not guarantee freedom from overfitting.

Question 19Introduction to Machine Learning

In a confusion matrix for a medical screening test, what does a false negative represent?

  1. A healthy patient incorrectly classified as diseased
  2. A diseased patient correctly classified as diseased
  3. A diseased patient incorrectly classified as healthy Correct answer.
  4. A healthy patient correctly classified as healthy

Explanation

A false negative occurs when the model predicts the negative class but the true label is positive, meaning a diseased patient is missed. This is often the most costly error in screening because disease goes undetected. A false positive is the reverse error.

Question 20Introduction to Machine Learning

Which statement about the relationship between model complexity and error is generally correct?

  1. Increasing complexity reduces training error but can increase test error beyond a point Correct answer.
  2. Model complexity has no effect on either training or test error
  3. Increasing complexity always reduces test error
  4. Decreasing complexity always reduces test error

Explanation

As model complexity grows, training error generally falls because the model can fit more detail. Test error initially falls but then rises once the model starts fitting noise, producing the classic overfitting curve. The optimal complexity balances the two.

Question 21Introduction to Machine Learning

A model trained on data from one hospital is deployed in a hospital serving a very different population and performs poorly. What is the most likely explanation?

  1. The model was evaluated using cross-validation
  2. The model has a distribution shift between training and deployment data Correct answer.
  3. The model's loss function was minimised too effectively
  4. The training set contained too few features to compute the loss

Explanation

When deployment data differ systematically from training data, the learned mapping no longer reflects the new population, a problem known as distribution shift or dataset shift. Good training performance does not transfer across distributions. Addressing it requires representative data or adaptation.

Question 22Introduction to Machine Learning

Which scenario is the clearest example of data leakage?

  1. Splitting data into training and test sets before any preprocessing
  2. Scaling features using statistics computed from the entire dataset before splitting Correct answer.
  3. Removing duplicate rows that appear in both training and test sets
  4. Using cross-validation to estimate model performance

Explanation

Computing scaling statistics on the full dataset lets information from the test set influence training, which is data leakage and inflates reported performance. Preprocessing should be fitted on training data only and then applied to test data. Leakage produces optimistic estimates that fail in deployment.

Question 23Introduction to Machine Learning

A team must choose between a highly accurate but opaque model and a slightly less accurate interpretable model for credit decisions. Which consideration is most relevant?

  1. Opaque models can never be deployed in any industry
  2. Interpretable models are guaranteed to generalise better
  3. Regulatory requirements may demand explanations for individual decisions, favouring interpretability Correct answer.
  4. Accuracy is always the only criterion that matters in production systems

Explanation

In regulated domains such as credit, decisions often must be explainable to applicants and auditors, so interpretability can outweigh a small accuracy gain. This is a governance and ethics consideration, not merely a technical one. The trade-off should be assessed against legal and business requirements.

Question 24Introduction to Machine Learning

Why can a high R-squared value on training data fail to indicate a good regression model?

  1. R-squared only applies to classification problems and is meaningless for regression
  2. R-squared measures training time rather than fit
  3. R-squared on training data can be inflated by overfitting and says nothing about unseen data Correct answer.
  4. R-squared is always negative for well-fitted models

Explanation

R-squared computed on the training set reflects how well the model fits the data it has seen, which can be inflated when the model is overly complex. It does not measure performance on new data. Evaluation on held-out data is required to judge generalisation.

Question 25Introduction to Machine Learning

What distinguishes semi-supervised learning from purely supervised learning?

  1. It learns only from rewards received from an environment
  2. It uses a small labelled set together with a larger unlabelled set during training Correct answer.
  3. It always produces clustering rather than prediction
  4. It requires no data at all and relies on expert rules

Explanation

Semi-supervised learning combines a limited amount of labelled data with abundant unlabelled data, exploiting structure in the unlabelled portion to improve learning. This is valuable when labelling is expensive but raw data are plentiful. It differs from both fully supervised and purely unsupervised settings.

Question 26Introduction to Machine Learning

In the standard taxonomy of machine learning, which pair of learning types is distinguished primarily by whether the training data includes labelled target values?

  1. Parametric and non-parametric learning
  2. Instance-based and model-based learning
  3. Supervised and unsupervised learning Correct answer.
  4. Batch and online learning

Explanation

Supervised learning trains on examples that pair inputs with known target labels, while unsupervised learning trains on inputs alone and seeks structure such as clusters or latent factors. The other pairs describe how models are updated, how they generalise, or how many parameters they assume, not the presence of labels.

Question 27Introduction to Machine Learning

A model is trained to assign each news article to one of four editorial sections. Which output type does this task produce?

  1. A continuous numeric value
  2. One category from a fixed set of four Correct answer.
  3. An ordered ranking of all four sections
  4. A probability distribution over an unbounded label space

Explanation

Assigning an item to one of a fixed, mutually exclusive set of categories is multi-class classification, so the output is a single class label from that set. Regression would output a continuous value, ranking would order items, and an unbounded label space is not the case here.

Question 28Introduction to Machine Learning

Which statement best describes the relationship between artificial intelligence, machine learning and deep learning?

  1. Deep learning is the broadest field, artificial intelligence is a subset of it, and machine learning is a subset of artificial intelligence.
  2. The three terms are synonyms and can be used interchangeably in professional practice.
  3. Machine learning is the broadest field, artificial intelligence is a subset of it, and deep learning is a subset of artificial intelligence.
  4. Artificial intelligence is the broadest field, machine learning is a subset of it, and deep learning is a subset of machine learning. Correct answer.

Explanation

Artificial intelligence is the overarching goal of building systems that perform tasks requiring intelligence; machine learning is one approach within it that learns from data; deep learning is a family of machine learning using multi-layer neural networks. The nesting is therefore AI contains ML contains DL.

Question 29Introduction to Machine Learning

Which statement best captures what is meant by the term 'model' in machine learning?

  1. The evaluation metric reported after training
  2. A mathematical function with parameters fitted to data that maps inputs to predictions Correct answer.
  3. The hardware accelerator used to speed up computation
  4. The raw dataset used to train an algorithm

Explanation

A model is the learned mapping from input features to output predictions, usually expressed as a function whose parameters are estimated from training data. The dataset, hardware and metrics are separate artefacts that support training and assessment but are not the model itself.

Question 30Introduction to Machine Learning

Which statement best describes the role of features in a machine learning model?

  1. Features are the target values the model is trained to predict.
  2. Features are the measurable input variables the model uses to make predictions. Correct answer.
  3. Features are the hyperparameters chosen before training begins.
  4. Features are the performance scores computed on the test set.

Explanation

Features, also called predictors or attributes, are the input variables that describe each example and are used by the model to produce an output. Targets are what is predicted, hyperparameters configure training, and scores measure results.

Question 31Introduction to Machine Learning

Which statement best describes the purpose of a validation set during model development?

  1. To store the model's learned parameters after training completes
  2. To provide the final unbiased estimate of performance on unseen data
  3. To increase the number of training examples available to the algorithm
  4. To tune hyperparameters and compare candidate models without touching the test set Correct answer.

Explanation

A validation set is held out from training and used to select hyperparameters and compare model variants, so that the test set remains untouched for a final unbiased estimate. It does not add training data, and it is not where parameters are stored.

Question 32Introduction to Machine Learning

Which statement best explains why a separate test set is required when reporting the performance of a machine learning model?

  1. It removes the need for a validation set in every workflow.
  2. It guarantees that the model will achieve the same accuracy in production.
  3. It increases the number of features available to the model.
  4. It provides an estimate of generalisation to data the model has not seen during training or model selection. Correct answer.

Explanation

The test set simulates unseen data and gives an estimate of how the model generalises beyond the examples used to fit and tune it. It cannot guarantee production performance, does not eliminate the need for validation during development, and does not change the feature set.

Question 33Introduction to Machine Learning

Which statement best describes the difference between parameters and hyperparameters in machine learning?

  1. Parameters apply only to neural networks, while hyperparameters apply only to decision trees.
  2. Parameters are learned from data during training, while hyperparameters are set before training to control the learning process. Correct answer.
  3. Parameters are chosen by the data scientist before training, while hyperparameters are learned from data.
  4. Parameters and hyperparameters are both learned automatically and differ only in name.

Explanation

Model parameters, such as weights or split thresholds, are estimated from the training data by the learning algorithm. Hyperparameters, such as learning rate or maximum tree depth, are configured beforehand and govern how training proceeds.

Question 34Introduction to Machine Learning

A hospital wants to estimate the likely length of stay in days for incoming patients based on admission data. Which machine learning task does this represent?

  1. Binary classification
  2. Dimensionality reduction
  3. Clustering
  4. Regression Correct answer.

Explanation

Length of stay measured in days is a continuous quantity, so predicting it is a regression task. Classification would apply to discrete categories, while clustering and dimensionality reduction are unsupervised techniques that do not predict a numeric target.

Question 35Introduction to Machine Learning

A retailer wants to group customers into segments based on purchase behaviour without using any predefined segment labels. Which learning approach is appropriate?

  1. Unsupervised clustering Correct answer.
  2. Supervised regression
  3. Supervised classification
  4. Reinforcement learning

Explanation

Because no segment labels are available, the task is to discover structure in the data, which is unsupervised learning; clustering is the standard method for forming such groups. Supervised methods require labels, and reinforcement learning learns from reward signals through interaction.

Question 36Introduction to Machine Learning

Which statement best describes why the bias-variance trade-off matters when selecting a machine learning model?

  1. It proves that the most complex model always has the lowest test error.
  2. It explains how model complexity affects both systematic error and sensitivity to training data, guiding the choice of complexity. Correct answer.
  3. It guarantees that a model with low bias will also have low variance.
  4. It shows that increasing training data always reduces variance to zero.

Explanation

Bias arises from overly simple assumptions and variance from sensitivity to the particular training sample; increasing complexity usually lowers bias but raises variance. The trade-off helps practitioners choose a complexity that balances the two rather than assuming one extreme is best.

Question 37Introduction to Machine Learning

A team is building a system that must decide, step by step, how to balance load across servers and receives a reward for low latency. Which learning paradigm fits this problem?

  1. Unsupervised learning with clustering
  2. Supervised learning with labelled optimal actions
  3. Dimensionality reduction
  4. Reinforcement learning through interaction and reward Correct answer.

Explanation

The system learns a policy by taking actions in an environment and receiving rewards, which is the defining setup of reinforcement learning. Labelled optimal actions are typically unavailable, and clustering or dimensionality reduction do not learn sequential decision policies.

Question 38Introduction to Machine Learning

Which statement best describes the purpose of feature scaling before training a distance-based model such as k-nearest neighbours?

  1. It ensures that features measured on larger numeric ranges do not dominate the distance calculation. Correct answer.
  2. It removes the need to split the data into training and test sets.
  3. It guarantees that the model will not overfit.
  4. It converts a regression problem into a classification problem.

Explanation

Distance-based methods compute similarities across features, so a feature with a large numeric range can overwhelm others unless features are rescaled to comparable ranges. Scaling does not change the task type, remove the need for data splits, or guarantee freedom from overfitting.

Question 39Introduction to Machine Learning

Which statement best describes the role of a loss function during training?

  1. It reports the final accuracy on the test set.
  2. It stores the training data for later reuse.
  3. It defines the quantity that the optimiser seeks to minimise by adjusting model parameters. Correct answer.
  4. It selects which features are used as inputs to the model.

Explanation

A loss function measures the discrepancy between predictions and targets, and training adjusts parameters to reduce this quantity. Feature selection, data storage and final test reporting are separate concerns.

Question 40Introduction to Machine Learning

A model achieves high accuracy on the training data but much lower accuracy on the test data. Which conclusion is most appropriate?

  1. The model is underfitting and should be made more complex.
  2. The test set is invalid and should be discarded.
  3. The training data contains no useful signal.
  4. The model is overfitting and may need regularisation or more training data. Correct answer.

Explanation

A large gap between training and test performance is the classic sign of overfitting, where the model has captured noise specific to the training sample. Remedies include regularisation, simpler models or more data, rather than discarding the test set.

Question 41Introduction to Machine Learning

Which statement best explains why a confusion matrix is more informative than accuracy alone for an imbalanced classification problem?

  1. It computes the exact probability that the model is correct.
  2. It guarantees that precision and recall are equal.
  3. It shows the counts of true positives, true negatives, false positives and false negatives, revealing per-class errors that accuracy can hide. Correct answer.
  4. It replaces the need for a test set.

Explanation

Accuracy aggregates all predictions into one number and can look high when the majority class dominates. A confusion matrix breaks performance down by outcome, exposing the specific error types that matter in imbalanced settings.

Question 42Introduction to Machine Learning

Which statement best describes the purpose of regularisation in machine learning?

  1. To increase the number of features used by the model
  2. To remove the need for a validation set
  3. To penalise model complexity and reduce overfitting Correct answer.
  4. To guarantee that training error reaches zero

Explanation

Regularisation adds a penalty for large or complex parameter values, discouraging the model from fitting noise and improving generalisation. It does not add features, replace validation, or aim to drive training error to zero.

Question 43Introduction to Machine Learning

A data scientist standardises features using the mean and standard deviation of the full dataset before splitting into training and test sets. Which problem does this introduce?

  1. Underfitting, because standardisation simplifies the model
  2. Data leakage, because information from the test set influences preprocessing Correct answer.
  3. An increase in the number of classes
  4. A reduction in the number of usable features

Explanation

Computing scaling statistics on the full dataset allows information from the test set to influence training, which is a form of data leakage and inflates the apparent performance. The correct practice is to fit the scaler on the training set only and apply it to the test set.

Question 44Introduction to Machine Learning

Which statement best explains why cross-validation is often preferred over a single train-validation split when data are limited?

  1. It reuses the data for multiple train-validation folds, giving a more stable performance estimate. Correct answer.
  2. It eliminates the need for a test set entirely.
  3. It guarantees that the final model will have no variance.
  4. It automatically selects the best features for the model.

Explanation

Cross-validation rotates which portion of the data is held out, so every example serves in validation once and performance is averaged over folds, reducing the dependence on a single arbitrary split. It does not remove variance, replace the test set, or perform feature selection by itself.

Question 45Introduction to Machine Learning

Which statement best describes the difference between batch learning and online learning?

  1. Batch learning is unsupervised, while online learning is supervised.
  2. Batch learning trains on the full dataset at once, while online learning updates the model incrementally as new data arrive. Correct answer.
  3. Batch learning always produces deeper neural networks than online learning.
  4. Batch learning requires labels, while online learning does not.

Explanation

Batch learning fits the model using all available training data in one or more passes, whereas online learning processes data sequentially and updates the model as each example or mini-batch arrives. Both can be supervised or unsupervised, and neither is defined by network depth.

Question 46Introduction to Machine Learning

A team has a small labelled dataset and a large collection of unlabelled records. Which approach is most appropriate for improving the model while exploiting the unlabelled data?

  1. Applying reinforcement learning with no reward signal
  2. Semi-supervised learning that combines the labelled examples with the unlabelled records Correct answer.
  3. Treating the unlabelled records as if they had positive labels
  4. Discarding the unlabelled records and training only on the labelled set

Explanation

Semi-supervised learning is designed for exactly this situation, using a small labelled set together with a large unlabelled set to learn better representations or decision boundaries. Discarding the unlabelled data wastes information, assigning false labels corrupts training, and reinforcement learning without rewards cannot learn.

Question 47Introduction to Machine Learning

A team must choose between a complex ensemble with 92% test accuracy and a simple logistic regression with 89% test accuracy for a regulated decision process that requires explanations. Which consideration is most relevant?

  1. The trade-off between predictive performance and interpretability required by the regulatory context Correct answer.
  2. The logistic regression should be chosen because simpler models always generalise better
  3. The choice should be made by comparing training accuracy only
  4. The ensemble should always be chosen because accuracy is the only criterion

Explanation

In regulated settings, the ability to explain decisions can be a hard requirement, so the team must weigh the accuracy gain of the ensemble against the interpretability of the simpler model. Accuracy alone is not the sole criterion, and training accuracy is not a reliable basis for model selection.

Question 48Introduction to Machine Learning

A model trained on data from one hospital performs poorly when deployed at a hospital serving a very different population. Which explanation is most consistent with this outcome?

  1. Distribution shift, because the deployment data differ from the training distribution Correct answer.
  2. Underfitting, because the model was too simple for the original hospital
  3. Overfitting to the test set, because the test set was too small
  4. Data leakage, because the original training data were labelled

Explanation

The model learned patterns specific to the original population, and the new population represents a different data distribution, so performance degrades. This is distribution shift rather than overfitting to a test set, underfitting on the original data, or leakage from correct labelling.

Question 49Introduction to Machine Learning

A team evaluates twenty model variants on the same validation set and selects the one with the best validation score. Which risk does this practice introduce?

  1. The selected model is guaranteed to be the best on unseen data
  2. The model's hyperparameters are no longer learned from data
  3. The validation score becomes optimistically biased because of repeated selection on the same data Correct answer.
  4. The training set becomes contaminated with test data

Explanation

Repeatedly comparing many variants on one validation set allows chance fluctuations to influence the selection, so the winning score overstates true performance. A separate test set or nested validation helps obtain an unbiased final estimate.

Question 50Introduction to Machine Learning

A dataset contains 990 healthy patients and 10 diseased patients. A model that always predicts 'healthy' reports 99% accuracy. Which statement best interprets this result?

  1. Accuracy is misleading here because the majority class dominates, so recall on the diseased class should also be examined Correct answer.
  2. The model has low precision on the healthy class
  3. The model is excellent because 99% accuracy is high by any standard
  4. The model has high recall on the diseased class

Explanation

With such imbalance, predicting the majority class yields high accuracy while detecting none of the diseased patients, so the accuracy figure hides a complete failure on the minority class. Metrics such as recall, precision and balanced accuracy on the diseased class are needed to reveal this.