A4.2 Data preprocessing (HL only)

The question asks about the effect of including a feature that has no plausible relationship with the target variable (readmission rate). This is the concept of an irrelevant feature.

An irrelevant feature is one with no meaningful statistical or causal relationship with the target variable. Wristband colour has no plausible connection to whether a patient is readmitted.

Including irrelevant features introduces noise into the model. The model may learn spurious correlations from the training data that do not generalise to new patients, increasing the risk of overfitting.

The correct answer is: It is an irrelevant feature that introduces noise and may reduce model generalisation. This directly reflects the definition and consequence of an irrelevant feature.

We need to identify what impact including wristband colour — a clearly unrelated attribute — would have on a readmission prediction model.

'More features always provide richer information' is incorrect. A well-established principle in machine learning is that adding irrelevant features degrades model performance by introducing noise, not improving it.

'Redundant because it is correlated with blood pressure' is incorrect. Wristband colour is arbitrarily assigned and has no statistical correlation with blood pressure. It is irrelevant, not redundant.

'Embedded methods will assign it zero' may be partially true in practice (e.g., Lasso could zero it out), but this does not describe the impact of including it — and relying on the model to fix poor data is not best practice.

The remaining option correctly identifies wristband colour as an irrelevant feature that introduces noise and risks reducing the model's ability to generalise to new data.

This question tests whether students can distinguish between two related but fundamentally different preprocessing techniques: feature selection and dimensionality reduction.

Feature selection chooses a subset of the original features to keep. The retained features remain in their original, interpretable form — for example, keeping 'Floor Area' and 'Bedrooms' from a housing dataset.

Dimensionality reduction (e.g., PCA) transforms the original features into a new set of components. These new components are mathematical combinations of the originals and are typically no longer directly interpretable.

The correct answer is: Feature selection retains a subset of the original features in their original form, while dimensionality reduction transforms features into new components. This is the precise distinction described in the study notes.

We must identify the correct distinction between feature selection and dimensionality reduction.

Option A reverses the definitions — it incorrectly states that feature selection transforms features and dimensionality reduction retains them. This is the opposite of the truth.

Option C is false — feature selection can be applied to both categorical and numerical features (e.g., chi-square for categorical, correlation coefficient for numerical). There is no data-type restriction on either technique.

Option D is incorrect — the two processes are fundamentally different in what they produce. Feature selection outputs a subset of original features; dimensionality reduction outputs new transformed components. They are not identical.

Option B is the only accurate description: feature selection retains original features in their interpretable form, while dimensionality reduction creates new transformed components.