machine learning chapter 2 & 3

The typical error size a model makes in predictions; penalizes large errors more heavily.

What does RMSE (Root Mean Square Error) measure?

When outliers are rare, since RMSE is more sensitive to large errors.

When is RMSE preferred over MAE?

The average magnitude of errors, treating all errors equally.

What does MAE (Mean Absolute Error) measure?

When the dataset has many outliers, since it’s less sensitive to them.

When is MAE preferred?

It ensures that each subgroup (stratum) of the population is fairly represented in both sets.

Why is stratified sampling important when splitting train/test sets?

How strongly two numerical features are linearly related (ranges from -1 to +1).

What does the Pearson Correlation Coefficient tell us?

Estimator: trained with .fit(); Transformer: has .transform(); Predictor: has .predict().

In Scikit-Learn, what are estimators, transformers, and predictors?

To chain preprocessing steps and model training into one workflow.

What is the purpose of a Scikit-Learn Pipeline?

A method to evaluate a model by splitting data into k folds and rotating which fold is the test set.

What is K-Fold Cross Validation?

Tests combinations of hyperparameters to find the best performing set.

What does Grid Search do in ML?

To show how many predictions were correct or incorrect across classes (TP, TN, FP, FN).

What is the purpose of a confusion matrix?

The model correctly predicts a positive instance.

What does a True Positive (TP) mean?

The model incorrectly predicts positive for a negative instance.

What does a False Positive (FP) mean?

Out of all predicted positives, how many were actually correct.

What does Precision measure?

Out of all actual positives, how many the model correctly identified.

What does Recall measure?

Increasing precision usually lowers recall, and increasing recall usually lowers precision.

What is the Precision–Recall trade-off?

The trade-off between the True Positive Rate (recall) and the False Positive Rate across thresholds.

What does the ROC curve show?

Predicting one label from more than two classes (e.g., digits 0–9).

What is multiclass classification?

One-vs-Rest (OvR) and One-vs-One (OvO).

What strategies are used for binary classifiers to handle multiclass?

Predicting multiple binary labels for one instance (e.g., 'movie is funny' and 'movie is romantic').

What is multilabel classification?

A generalization where each label can take multiple values (e.g., predicting multiple pixels with multiple intensity levels).

What is multioutput classification?