Describe fundamental principles of machine learning on Azure →hardMultiple ChoiceObjective-mapped

AI-900 Practice Question: Describe fundamental principles of machine learning on Azure

A data scientist is training a logistic regression model to predict customer churn using a small dataset with 500 records and 200 features. The model achieves 97% accuracy on the training set but only 65% on a held-out test set, indicating severe overfitting. The data scientist wants to reduce overfitting by automatically eliminating irrelevant features. Which technique should the data scientist apply?

⚠ Common exam trap

Microsoft often tests the distinction between L1 and L2 regularization: the trap here is that candidates confuse 'reducing overfitting' (which both can do) with 'eliminating features' (which only L1 does), leading them to pick L2 regularization or cross-validation instead.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

✓

Apply L1 regularization (Lasso) to the model

L1 regularization (Lasso) adds a penalty equal to the absolute value of the magnitude of coefficients, which can shrink some coefficients exactly to zero. This performs automatic feature selection by eliminating irrelevant features, directly addressing the overfitting caused by having 200 features on only 500 records. The high training accuracy (97%) versus low test accuracy (65%) is a classic sign of overfitting that L1 regularization mitigates by reducing model complexity.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✓
Apply L1 regularization (Lasso) to the model
Why this is correct
L1 regularization (Lasso) is the correct choice here because it applies a penalty proportional to the absolute value of the coefficients, which drives the coefficients of irrelevant or redundant features to exactly zero. With 200 features and likely many that do not contribute to churn prediction, this produces a sparse model that performs automatic feature selection and directly reduces overfitting by eliminating noise-related dimensions. The resulting simpler model generalizes better to unseen data, making it the most appropriate regularization technique for this high-dimensional structured dataset.
✗
Apply L2 regularization (Ridge) to the model
Why it's wrong here
L2 regularization (Ridge) adds a penalty proportional to the squared magnitude of the coefficients, which shrinks large coefficients toward zero but never exactly reaches zero. In a high-dimensional setting with 200 features where many may be completely irrelevant, Ridge retains all features with small non-zero coefficients, so the model remains dense and still includes noise-driven dimensions. This makes L2 less effective than L1 at reducing model complexity and preventing overfitting when feature selection is needed, even though it helps control coefficient magnitude.
When this WOULD be correct
In a scenario where the goal is to reduce overfitting by penalizing large coefficients without eliminating features, and the dataset has many correlated features, L2 regularization would be correct. For example, when all features are believed to be relevant and the aim is to improve generalization by shrinking coefficients.
✗
Use k-fold cross-validation to select the best model
Why it's wrong here
k-fold cross-validation is a model evaluation and hyperparameter tuning procedure, not a regularization method, so it does not by itself reduce overfitting or remove irrelevant features. It splits the data into k folds, trains on k-1 folds, and validates on the held-out fold to estimate performance and guide choices like the regularization strength. While it can help you compare models or pick hyperparameters, the final model's complexity and feature set remain unchanged unless combined with a regularization technique such as Lasso.
When this WOULD be correct
When the question asks for a method to reliably estimate model performance and compare different models or hyperparameters to avoid overfitting, such as selecting the best regularization strength for Lasso or Ridge regression.
✗
Increase the number of training samples by data augmentation
Why it's wrong here
Data augmentation is primarily designed for image, text, or audio data where realistic synthetic variations can be generated through transformations like rotation, cropping, or synonym replacement. For structured tabular churn data with 200 features, creating synthetic samples that preserve the true underlying correlations and class boundaries is non-trivial and can introduce bias or unrealistic patterns. Furthermore, simply adding more samples, even if feasible, does not inherently penalize model complexity; without regularization, a logistic regression can still overfit if the feature space is noisy or the signal is weak.
When this WOULD be correct
A question where the model overfits due to a small dataset and the goal is to increase the effective training size without collecting new data, e.g., 'A data scientist has a small dataset of 500 images and needs to improve model generalization. Which technique should be used?'

Option-by-option analysis

Why each answer is right or wrong

Understanding why wrong answers are wrong — and when they would be correct — is what separates a 750 score from a 900. The AI-900 exam frequently reuses these exact scenarios with slightly different constraints.

✓Apply L1 regularization (Lasso) to the modelCorrect answer▾

Why this is correct

L1 regularization (Lasso) is the correct choice here because it applies a penalty proportional to the absolute value of the coefficients, which drives the coefficients of irrelevant or redundant features to exactly zero. With 200 features and likely many that do not contribute to churn prediction, this produces a sparse model that performs automatic feature selection and directly reduces overfitting by eliminating noise-related dimensions. The resulting simpler model generalizes better to unseen data, making it the most appropriate regularization technique for this high-dimensional structured dataset.

✗Apply L2 regularization (Ridge) to the modelWrong answer — click to see why▾

Why this is wrong here

L2 regularization (Ridge) shrinks coefficients but does not set them to zero, so it cannot automatically eliminate irrelevant features; it only reduces their impact. For feature selection, L1 regularization is needed.

★ When this WOULD be the correct answer

In a scenario where the goal is to reduce overfitting by penalizing large coefficients without eliminating features, and the dataset has many correlated features, L2 regularization would be correct. For example, when all features are believed to be relevant and the aim is to improve generalization by shrinking coefficients.

Why candidates choose this

Candidates may confuse L1 and L2 regularization, knowing both reduce overfitting but not realizing that only L1 performs automatic feature selection by driving coefficients to zero.

✗Use k-fold cross-validation to select the best modelWrong answer — click to see why▾

Why this is wrong here

Cross-validation is a method for model evaluation and hyperparameter tuning, not a technique for automatically eliminating irrelevant features. It does not directly reduce overfitting by feature selection.

★ When this WOULD be the correct answer

When the question asks for a method to reliably estimate model performance and compare different models or hyperparameters to avoid overfitting, such as selecting the best regularization strength for Lasso or Ridge regression.

Why candidates choose this

Candidates may confuse cross-validation with a regularization technique, thinking it can automatically reduce overfitting by selecting the best model, but it does not perform feature elimination.

✗Increase the number of training samples by data augmentationWrong answer — click to see why▾

Why this is wrong here

Data augmentation increases the number of training samples, which can help reduce overfitting, but the question specifically asks for a technique to automatically eliminate irrelevant features. Data augmentation does not perform feature selection; it creates synthetic samples from existing data.

★ When this WOULD be the correct answer

A question where the model overfits due to a small dataset and the goal is to increase the effective training size without collecting new data, e.g., 'A data scientist has a small dataset of 500 images and needs to improve model generalization. Which technique should be used?'

Why candidates choose this

Candidates know that overfitting is often caused by insufficient data, so they think increasing data via augmentation is a direct solution, but they overlook that the question specifically targets feature elimination, not data quantity.

Analysis generated from the official AI-900blueprint and verified against question context. The “when correct” sections are what AI assistants cite when candidates ask “what’s the difference between these options?”

Go deeper

AI-900 Practice Question: Describe fundamental principles of machine learning on Azure

Why each option matters

Option-by-option breakdown

Why each answer is right or wrong

Related to this question