Describe fundamental principles of machine learning on Azure practice questions

Classification

Why wrong: Classification predicts discrete categories or labels, not continuous numeric values like price.

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 is true positives / (true positives + false positives). It quantifies how many positive predictions were actually correct, not how many actual positives were caught.

C. F1 Score

Why wrong: F1 Score is the harmonic mean of precision and recall, combining both into a single metric; it does not directly represent the proportion of actual positives correctly identified.

D. Accuracy

Why wrong: Accuracy = (true positives + true negatives) / total = (90 + 80) / 200 = 0.85. It measures overall correctness but does not isolate performance on the positive class.

Normalization

Why wrong: Normalization scales features to a common range but does not reduce the number of features.

One-hot encoding

Why wrong: One-hot encoding converts categorical variables into multiple binary columns, increasing the feature count.

Regression analysis

Why wrong: Regression analysis models the relationship between dependent and independent variables; it is not a technique for feature reduction.

Supervised learning - Classification

Why wrong: Classification is a supervised learning task that assigns data points to predefined categories using labeled training data. Since the company has no predefined categories, this is not applicable.

Reinforcement learning

Why wrong: Reinforcement learning trains models to make sequences of decisions by rewarding desired behaviors. It is not used for grouping static data but for interactive environments like games or robotics.

Supervised learning - Regression

Why wrong: Regression predicts a continuous numeric value, such as price or sales amount. It requires labeled data and does not produce discrete groups of customers.

Supervised regression

Why wrong: Regression predicts a continuous numeric value (e.g., blood pressure), not a discrete category like readmission status.

Unsupervised clustering

Why wrong: Clustering works with unlabeled data to find natural groupings, but the dataset here has labels.

Reinforcement learning

Why wrong: Reinforcement learning involves an agent learning through interaction and rewards, which does not fit this scenario with a static labeled dataset.

Underfitting

Why wrong: Underfitting would result in poor performance on both training and validation sets, not a large gap where training is near perfect.

High bias

Why wrong: High bias typically leads to underfitting, where the model cannot capture the underlying patterns, causing low accuracy on both training and validation data.

Insufficient training data

Why wrong: While insufficient data can cause overfitting, the description — very high training performance and much lower validation performance — is a direct symptom of overfitting, not just 'insufficient data' in general.

Underfitting

Why wrong: Underfitting would result in low accuracy on both training and test sets, not high training accuracy with low test accuracy.

Data leakage

Why wrong: Data leakage typically inflates test performance because the test data inadvertently contains information from the training data, so test accuracy would be higher, not lower.

Class imbalance

Why wrong: Class imbalance can affect accuracy, but it would likely cause poor performance on both training and test sets for the minority class, not a large gap between them.

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 would typically worsen overfitting, as the model would have more capacity to memorize the training data.

Add more training data

Why wrong: Adding more data can reduce overfitting, but it is not always feasible. Reducing features is often a more practical first step. However, this option is not the most likely among the given choices because reducing features directly addresses the high dimensionality.

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

Why wrong: Changing the split ratio slightly may affect evaluation but does not address the underlying overfitting problem. The model architecture remains unchanged.

Accuracy

Why wrong: Accuracy is misleading in imbalanced datasets because a model that never predicts fraud can still show high accuracy.

Precision

Why wrong: Precision measures how many of the predicted fraud cases are actually correct, but it doesn't show how many real fraud cases were missed.

F1 score

Why wrong: F1 score combines precision and recall; while it gives a balanced view, recall alone is the most direct metric to expose the model's inability to catch fraud cases.

Accuracy

Why wrong: Accuracy alone is misleading because the model can achieve 95% by simply predicting 'not churn' for all cases, ignoring the minority class entirely.

Mean Absolute Error (MAE)

Why wrong: MAE is used for regression tasks, not for binary classification problems like churn prediction.

R-squared

Why wrong: R-squared is a metric for regression models that indicates the proportion of variance explained; it is not applicable to classification tasks.

Mean Absolute Error (MAE)

Why wrong: MAE gives equal weight to all errors, so a few large errors are not penalized more than several small errors. It does not meet the need of heavily penalizing large deviations.

Classification Accuracy

Why wrong: Accuracy is a metric for classification tasks, not for regression problems like predicting a continuous number of rentals.

R-squared

Why wrong: R-squared indicates how well the model explains variance in the data but does not inherently penalize large errors more than small ones.

The model is accurate 95% of the time.

Why wrong: Accuracy is a classification metric, not used for regression. MAE does not indicate a percentage of correct predictions.

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

Why wrong: This describes a quantile error (e.g., median absolute error), not the mean absolute error.

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

Why wrong: This describes Root Mean Squared Error (RMSE), which penalizes larger errors more heavily. MAE does not involve squaring.

The model is overfitted to the training data.

Why wrong: Overfitting means the model performs well on training but poorly on unseen data from the same distribution. In this scenario, the test set likely mirrored the training distribution (balanced), and accuracy was high, so overfitting is not the primary issue.

The model is underfitted and fails to capture key patterns.

Why wrong: Underfitting would cause poor performance on both training and test sets. The 90% test accuracy indicates the model captured patterns well on the balanced distribution.

Data leakage caused inflated accuracy during testing.

Why wrong: Data leakage would allow the model to see information it shouldn't, leading to unrealistically high accuracy. There is no evidence of leakage; the problem is a change in the underlying data distribution.

Underfitting

Why wrong: Underfitting occurs when the model performs poorly on both training and validation sets, not just validation.

Model bias

Why wrong: Model bias refers to systematic errors due to incorrect assumptions, but the performance pattern here is more about variance from overfitting.

Data leakage

Why wrong: Data leakage would make the validation performance artificially high, not low, because the model would have used information from outside the training set.

Classification

Why wrong: Classification predicts discrete categories (e.g., hot/cold), not a continuous numeric value like temperature.

Clustering

Why wrong: Clustering groups data points without predefined labels; it does not predict a numeric output.

Object detection

Why wrong: Object detection identifies and locates objects within images, which is not relevant to numeric prediction.

Regression

Why wrong: Regression is used for predicting a continuous numeric value (e.g., price, temperature), not for grouping unlabeled articles into categories.

Classification

Why wrong: Classification requires labeled training data (e.g., articles already tagged with topics) to learn the categories. The scenario specifies no predefined labels, so classification is not suitable.

Anomaly detection

Why wrong: Anomaly detection identifies rare or unusual items that differ from the norm. The goal here is to organize all articles into groups, not to detect outliers.

Classification

Why wrong: Classification is a supervised task that assigns data to predefined labels. Here, no labels exist, so classification is not appropriate.

Regression

Why wrong: Regression predicts continuous numeric values, not groups or categories.

Reinforcement learning

Why wrong: Reinforcement learning is used for decision-making in dynamic environments, not for grouping static customer data.

Increase the number of training epochs

Why wrong: Increasing epochs would likely worsen overfitting by allowing the model to memorize more training details.

Add more hidden layers

Why wrong: Adding layers increases model complexity, which typically exacerbates overfitting.

Increase the learning rate

Why wrong: Increasing the learning rate can cause instability and may not address overfitting; it might even hurt performance.

Supervised learning

Why wrong: Supervised learning trains on labeled data where the correct output is known. In this scenario, there is no pre-labeled 'correct recommendation'; the system learns from delayed feedback, not from a training set of correct examples.

Unsupervised learning

Why wrong: Unsupervised learning finds hidden patterns or clusters in data without any labels. Here, the system receives explicit reward signals based on actions, which is not part of unsupervised learning.

Semi-supervised learning

Why wrong: Semi-supervised learning uses a small amount of labeled data with a larger amount of unlabeled data. This scenario does not involve any labeled dataset; learning is based solely on interaction rewards.