Describe fundamental principles of machine learning on Azure practice questions

Increase the size of the training dataset

Why wrong: More data can help, but it is not a direct modeling technique; regularization is applied at training time to constrain the model and prevent memorization.

Increase the complexity of the model (e.g., add more features)

Why wrong: Increasing complexity typically makes overfitting worse, as the model becomes more capable of memorizing noise.

Switch to a different regression algorithm

Why wrong: Changing algorithms may help, but regularization is a direct and standard technique for reducing overfitting in regression models.

Increase the number of features

Why wrong: Adding more features increases model complexity, which often worsens overfitting.

Decrease the training data size

Why wrong: Having less training data generally increases the risk of overfitting, as the model has fewer examples to generalize from.

Increase the number of training epochs

Why wrong: More epochs can lead to further overfitting as the model becomes more specialized to the training data.

Accuracy

Why wrong: Accuracy is supported in Azure AutoML, but it can be misleading when class distribution changes; it does not explicitly guarantee equal per-class importance unless classes are perfectly balanced.

Weighted F1

Why wrong: Weighted F1 (f1_score_weighted) is a supported AutoML primary metric, but it weights classes by support/frequency, so it does not give equal importance to each species.

Micro F1

Why wrong: Micro F1 is not a selectable primary metric in Azure AutoML classification, and micro averaging pools contributions, which can favor classes with more examples.

Cross-validation

Why wrong: Cross-validation, especially k-fold, partitions the training data into multiple folds to train and validate the model repeatedly. This provides a robust estimate of the model's out-of-sample performance by averaging across different subsets, thereby reducing the impact of a particular data split. It helps diagnose and mitigate overfitting by exposing the model to varied customer segments during validation, ensuring hyperparameters are tuned for generalization rather than memorization.

Increasing model complexity

Why wrong: Increasing complexity, such as adding more layers or parameters, gives the model greater capacity to fit noise in the training data. This leads to low bias but high variance where the model captures idiosyncrasies of the training set that do not hold for new customer segments. Consequently, test performance degrades, and the model becomes less reliable for predicting churn across diverse populations.

Using a larger learning rate

Why wrong: A large learning rate amplifies parameter updates, causing the optimizer to overshoot the minima and potentially oscillate or diverge. Even if it converges, the final model may land in a high-loss region, yielding poor performance on both training and validation data. Furthermore, large learning rates can cause unstable training dynamics, making the model sensitive to small data perturbations and hurting generalization to new segments.

A. Precision for the default class

Why wrong: Precision is the fraction of positive predictions that are correct. Since the model makes no positive predictions, precision is undefined (division by zero) and not helpful.

C. F1-score for the default class

Why wrong: F1-score is the harmonic mean of precision and recall. Since recall is 0, the F1-score is also 0, but recall alone already reveals the issue more directly.

D. Overall accuracy

Why wrong: Accuracy is 95% because the model correctly predicts the majority class (non-default) for all cases. This high value masks the complete failure to detect defaults.

Overall accuracy

Why wrong: Overall accuracy is misleading here because it is calculated as the ratio of all correct predictions to the total number of predictions, which is dominated by the majority class in an imbalanced dataset. Even if the model misclassifies most Technical samples, those errors are numerically small compared to correct predictions on the majority classes, so accuracy remains high. This is the accuracy paradox: a high overall accuracy can hide severe underperformance on a rare but important class.

Confusion matrix

Why wrong: A confusion matrix is a valuable diagnostic table showing counts of true positives, false positives, true negatives, and false negatives for every pair of classes, but it is not a single metric. While it would visually reveal that Technical samples are being confused with other classes, the question specifically asks for a metric that best reveals the issue; relying on a raw matrix forces subjective interpretation and does not yield a concise, comparable number. F1-score condenses the matrix's precision and recall information into one clear score.

Precision for the 'General' class

Why wrong: Precision for the majority class does not reflect performance on the minority class. Precision for the 'Technical' class would be more relevant, but F1-score is even more informative as it accounts for recall as well.

Supervised learning requires a human to watch the training process; unsupervised runs automatically

Why wrong: This statement confuses operational monitoring with the technical definition of learning paradigms. Human oversight, such as monitoring for drift or retraining schedules, is an MLOps practice that applies to both supervised and unsupervised systems. The distinction between these paradigms is rooted purely in the data: supervised learning uses labeled examples, whereas unsupervised learning works with unlabeled data, and neither requires continuous human observation during training.

Supervised learning uses neural networks; unsupervised uses decision trees only

Why wrong: This statement falsely claims that algorithm families are intrinsically tied to learning paradigms. In practice, neural networks are widely used in unsupervised learning (e.g., autoencoders for representation learning), and decision trees are commonly employed in supervised learning (e.g., classification or regression trees). Both paradigms can leverage nearly any algorithm family—the real difference is solely whether training data includes labels or not.

Unsupervised learning always produces better results than supervised learning

Why wrong: This statement incorrectly asserts a universal performance hierarchy. Supervised learning excels in tasks where the correct answer is known, enabling precise evaluation and predictive accuracy; however, it cannot be used when labels are unavailable or expensive to obtain. Unsupervised learning offers insight into data structure but lacks ground truth for evaluating outputs, so neither paradigm is universally 'better'—their utility depends entirely on the problem context and data availability.

Classification

Why wrong: Predicting an exact market price is a regression task, not classification, because the target variable is a continuous numerical value rather than a discrete category. This option is tempting because classification is commonly used for predicting labels like “low”, “medium”, or “high” price brackets, which would be correct if the goal were to assign a house to a predefined price tier.

Clustering

Why wrong: Clustering groups unlabeled data into clusters based on similarity; it does not predict a specific numeric value.

Anomaly Detection

Why wrong: Anomaly Detection identifies data points that differ significantly from the majority, not suitable for predicting a continuous value.

A. Precision

Why wrong: Precision quantifies the reliability of positive predictions: it divides true positives by the sum of true positives and false positives, yielding 90/(90+20) = 0.818. It focuses on whether predicted spam is actually spam, not on whether actual spam was detected. The question asks about 'proportion of actual spam correctly identified,' which is the definition of recall, not precision.

C. F1 Score

Why wrong: F1 Score is the harmonic mean of precision and recall, and it collapses both metrics into a single composite number. Because it incorporates precision, it cannot isolate the proportion of actual positive cases that were caught; a model can have a high F1 score due to high precision even if its recall is lower. In this scenario, the F1 score would be approximately 0.857, not the direct 0.9 value needed.

D. Accuracy

Why wrong: Accuracy measures the ratio of all correct predictions (both true positives and true negatives) to the total number of cases, here (90+80)/200 = 0.85. It fails to address the question because it dilutes the model's performance on the positive class with the true negatives, which are irrelevant to identifying actual spam. Even if the model missed 10 spam emails, accuracy can appear high if the negative class is large.

Normalization

Why wrong: Normalization (e.g., min-max scaling or z-score standardization) adjusts the numeric range of feature values so that they share a common scale, such as 0 to 1 or zero mean and unit variance. This rescaling does not alter the number of features or their dimensionality; every original feature remains in the dataset, just with transformed values. Normalization is typically a preparatory step before applying PCA or distance-based algorithms, but by itself it does not address the need to reduce the feature set.

One-hot encoding

Why wrong: One-hot encoding is the opposite of dimensionality reduction: it converts a categorical variable with k categories into k binary indicator columns, thereby expanding the feature space. For retail data with many categorical attributes (e.g., product categories, payment methods), one-hot encoding can drastically increase the number of dimensions, leading to higher computational cost and a greater risk of the curse of dimensionality. It is a preprocessing transformation for handling categorical inputs, not a technique for reducing the number of features.

Regression analysis

Why wrong: Regression analysis instead models the statistical relationship between a dependent variable (e.g., likelihood to churn) and one or more independent variables by estimating coefficients. Although regularized forms like Lasso can implicitly perform feature selection by shrinking some coefficients to zero, the core technique does not reduce the number of features in the dataset; it requires the full set of predictors to fit the model. In the context of this question, regression is a supervised modeling approach, not a feature reduction method like PCA.

Supervised learning - Classification

Why wrong: Classification is a supervised learning task that assigns each data point to one of a predefined set of categories using a labeled training dataset with ground-truth labels. In this retail scenario, there are no existing customer segment labels or known group definitions; the company wants the model to identify segments from the data itself. Without labeled examples to learn from, classification cannot be used to automatically discover customer groups.

Reinforcement learning

Why wrong: Reinforcement learning trains an agent to make a sequence of decisions by interacting with an environment and maximizing a cumulative reward signal, commonly applied in robotics, game playing, or dynamic pricing. It is not designed for static grouping tasks because there is no state-transition process or reward for forming customer segments. Grouping customers into clusters is a one-time pattern discovery problem, not a sequential decision-making problem.

Supervised learning - Regression

Why wrong: Regression is a supervised learning technique that predicts a continuous numeric value, such as a customer's future spending or product price, by training on labeled datasets. It requires known target values to learn a mapping from features to a number, so it cannot output discrete group assignments or segment memberships. Because the company needs to create distinct customer groups rather than predict a numeric quantity, regression is not applicable to this scenario.

Supervised regression

Why wrong: Supervised regression is incorrect because the target variable is a discrete category, not a continuous numeric quantity. Regression would predict a value like length of hospital stay or blood pressure, whereas readmission status is a binary class. Even if a regressor could be thresholded into two groups, doing so would be poorly calibrated and would not optimize classification metrics such as F1-score, so it is not the right framework for this labeled categorical outcome.

Unsupervised clustering

Why wrong: Unsupervised clustering is wrong because it assumes the data has no ground-truth labels and seeks to discover hidden structures or patient subgroups on its own. This dataset does have labels (readmitted or not), so the learning problem is supervised. Using k-means or hierarchical clustering here would ignore those known outcomes and cannot directly output a model that predicts readmission status from a new patient's features.

Reinforcement learning

Why wrong: Reinforcement learning is not applicable because it trains an agent to take sequential actions in an environment and learn from delayed rewards or penalties. A historical patient dataset is static and passive: there is no action at each record, no environment state transition, and no reward signal based on choosing a policy. Readmission prediction is a one-shot pattern-recognition task, not an interactive decision-making process.

Underfitting

Why wrong: Underfitting occurs when a model is too simplistic to capture the underlying relationships in the data, resulting in low accuracy on both the training and validation sets. Instead of seeing a large gap with training near perfect, an underfit model would show similarly poor performance everywhere. Therefore, the scenario’s high training accuracy directly contradicts underfitting.

High bias

Why wrong: High bias means the model's assumptions are too rigid, so it systematically misses true patterns and produces high error irrespective of the dataset. This bias leads to underfitting, where even the training error remains high and can never approach near-perfect accuracy. Since the training performance is described as very high, high bias cannot be the cause.

Insufficient training data

Why wrong: Having too few training samples can certainly increase variance and make overfitting more likely, but it is a contributing factor rather than the observed behavior. The evidence—extremely high training performance and much lower validation performance—is the direct symptom of overfitting, and this gap can appear even with a large dataset if the model is too complex or trained too long. Thus, insufficient data is not the best answer because the described pattern is specifically overfitting's signature.

Underfitting

Why wrong: Underfitting is the opposite failure mode: the model is too simple to capture the underlying relationship (high bias), so it cannot achieve high accuracy even on the training set itself. Because the model has not learned the relevant patterns, its predictions are similarly poor on both training and test partitions, producing low accuracy everywhere without a widening generalization gap. Therefore a huge gap with strong training performance cannot be explained by underfitting.

Data leakage

Why wrong: Data leakage is a data collection or preprocessing flaw in which information from the test set (or the target variable, or future data) is inadvertently included when fitting the model. Leakage lets the model exploit patterns that will be artificially present in the test features too, which inflates test accuracy—the opposite of the reported low test accuracy. Because the leaked information also helps during training, it can make both training and test look good, so it cannot be the cause of a large drop from training to test.

Class imbalance

Why wrong: Class imbalance can degrade performance by causing the model to heavily favor the majority class, especially when accuracy is used as the metric, but it does not directly create a training-versus-test generalization gap. In a skewed dataset, a model may achieve decent overall accuracy by predicting the majority class everywhere, yet this pattern holds similarly on both training and test sets; the minority class suffers poor recall in both. Unless the test set's class distribution differs radically from the training set, imbalance alone would not produce high training accuracy with sharply lower test accuracy.

Binary classification

Why wrong: Binary classification predicts one of two categories (e.g., yes/no), not a continuous value.

Multiclass classification

Why wrong: Multiclass classification predicts one of multiple categories (e.g., animal types), not a continuous value.

Clustering

Why wrong: Clustering is an unsupervised learning task that groups similar data points without a predefined target variable.

Increase the complexity of the model by adding more layers or parameters

Why wrong: Increasing model complexity by adding more layers or parameters gives the model greater capacity to fit the training data almost perfectly, which amplifies overfitting. With a small dataset of 100 samples, an overly complex model will memorize idiosyncrasies and random noise, driving test error even higher. This is the opposite of the needed intervention and would worsen generalization rather than improve it.

Add more training data

Why wrong: Adding more training data is unlikely to resolve the high test error here because the model already exhibits overfitting—low training error paired with high test error—indicating it has memorised noise rather than learned generalisable patterns. This approach is tempting because additional data often reduces variance in underfitting scenarios where the model lacks sufficient examples to capture underlying trends. However, in this case, the core issue is excessive model complexity relative to the dataset, which requires regularisation or feature reduction, not merely more samples.

Use a different train-test split ratio like 80-20 instead of 70-30

Why wrong: Changing the train-test split ratio, such as moving from 70-30 to 80-20, only alters the proportion of data used for training versus validation, not the model's underlying architecture or complexity. Since the model remains equally over-parameterized, it will still overfit the training data, and any change in test error arises from different random sampling of a small dataset, not from fixing the variance problem. This approach does not address the root cause of overfitting.

Accuracy

Why wrong: Accuracy is the ratio of all correct predictions (true positives plus true negatives) to total predictions, and it becomes dangerously misleading when classes are imbalanced. With a fraud rate of just 1%, a model that classifies every transaction as legitimate still achieves 99% accuracy, yet catches none of the fraud. Because accuracy credits correct predictions on the dominant class, it can hide a complete lack of fraud detection and should not be used to judge such a model.

Precision

Why wrong: Precision is calculated as true positives divided by true positives plus false positives (TP/(TP+FP)), meaning it measures the purity of the fraud claims: how many flagged frauds are actually fraudulent. A model can achieve very high precision by flagging only a few obvious fraud cases while ignoring the majority of fraudulent transactions, since missed frauds (false negatives) do not affect precision. Thus precision tells you about the trustworthiness of the alerts, but it cannot show how many real fraud cases the model failed to catch.

F1 score

Why wrong: The F1 score is the harmonic mean of precision and recall, giving equal weight to both in a single number, but this balance often obscures the exact issue at hand. In fraud detection, a model with high precision and only moderate recall can have a reasonably high F1 score even though it is still missing many fraudulent transactions. Because only recall measures the false-negative rate directly, relying on F1 can mask the model's inability to catch fraud; the most direct signal of detection failure is recall.

Accuracy

Why wrong: Accuracy alone is misleading for imbalanced churn prediction because it only counts the percentage of correct predictions overall, including the majority class. If 95% of customers do not churn, a model that always predicts 'not churn' will achieve 95% accuracy while completely failing to identify any actual churners, which is exactly the group of interest. In churn prediction, the cost of missing a true churner (lost revenue, customer acquisition costs) is high, and accuracy does not reflect that trade-off. A high accuracy number can coexist with a recall of zero, so accuracy should be paired with precision, recall, or F1 to provide a realistic evaluation.

Mean Absolute Error (MAE)

Why wrong: Mean Absolute Error (MAE) is a regression metric that computes the average absolute difference between continuous predicted values and actual numeric targets. For a binary classification problem like churn prediction, the target is a discrete categorical label (e.g., churned or not churned), so there is no meaningful numeric distance between a predicted class and the true class. Even if the model outputs class probabilities rather than hard labels, MAE on probabilities would not account for the decision threshold or the class imbalance that is central to classification evaluation. Thus MAE is fundamentally designed for regression tasks and cannot assess classification performance.

R-squared

Why wrong: R-squared, also known as the coefficient of determination, measures the proportion of variance in a continuous target variable that is explained by the regression model. It relies on the concept of total sum of squares relative to the residual sum of squares, which only makes sense when predictions and actual values are numeric and a mean reference is meaningful. In churn classification, the outcomes are categorical and not measured on an interval scale, so the variance decomposition that R-squared performs has no logical equivalent. Therefore R-squared is inapplicable to classification models, regardless of how well the model predicts churn.

Mean Absolute Error (MAE)

Why wrong: Mean Absolute Error sums the absolute differences between predicted and actual rental counts, so each error contributes linearly and equally to the average. Consequently, one wildly inaccurate prediction is penalized no more than a collection of small errors, which directly contradicts the requirement to heavily avoid large deviations. Although MAE is robust to outliers and interpretable, it is not the right metric when large errors carry disproportionate business cost.

Classification Accuracy

Why wrong: Classification Accuracy computes the fraction of predictions whose labels exactly match the true class, which is meaningless for a continuous rental-count target. Rental counts are numerical values, not discrete categories, so there is no class label to match; using this metric would require artificially binning the counts and would ignore how far off a prediction is from the actual number. Accuracy is therefore fundamentally a classification metric, not a regression evaluation metric.

R-squared

Why wrong: R-squared measures the proportion of variance in observed rental counts explained by the model, producing a unitless value between 0 and 1. It is a relative goodness-of-fit score that depends on the total variance of the data, so it does not directly quantify or control the size of individual large residuals. A model with high R-squared can still have a few badly inaccurate predictions, making it unsuitable for a requirement to aggressively penalize large errors.

The model is accurate 95% of the time.

Why wrong: Accuracy is a classification metric that measures the proportion of correct predictions out of total predictions, which is not meaningful for regression tasks where predictions are continuous values. MAE does not report a percentage of correct predictions; it reports an average error magnitude in the same units as the target variable (dollars). Saying the model is 'accurate 95% of the time' would require defining a tolerance band for what counts as correct, but MAE makes no such threshold assumption.

The model's predictions are within $5,000 of the actual prices for 50% of the houses.

Why wrong: This option describes a percentile-based error statistic, such as the median absolute error, which reports the value below which 50% of the individual absolute errors fall. MAE, by contrast, is the arithmetic mean of all absolute errors and is sensitive to the presence of a few very large errors that pull the average upward. Knowing that half the predictions are within $5,000 does not imply the average error is $5,000; for instance, large errors on the other half could make the MAE much higher.

The square root of the average squared error is $5,000.

Why wrong: This is the definition of Root Mean Squared Error (RMSE), not Mean Absolute Error (MAE). RMSE calculates the square root of the mean of the squared differences between predictions and actual values, which gives disproportionate weight to large errors. MAE simply averages the absolute differences without squaring, so it treats all error magnitudes linearly and will always be less than or equal to RMSE for the same set of predictions.