1 Foundations preprocessing

How do classification, regression, clustering and association mining differ?

Classification predicts a categorical label; regression predicts a numerical response; clustering groups observations by similarity; association mining finds co-occurring item combinations.

What separates supervised from unsupervised learning?

Supervised training uses known target responses. Unsupervised methods discover structure without using a target label to fit that structure.

What is the difference between an observation and a feature?

An observation is the unit represented by one row; a feature is a measured attribute represented by a column. Their meaning depends on the analysis question.

What is the difference between a population and a sample?

The population is the collection about which conclusions are intended; a sample is the subset actually observed. A large sample can still be unrepresentative.

Why is an association not automatically a causal effect?

Co-occurrence can reflect shared causes, subgroup composition or selection. Manipulating one associated variable need not change the other.

What distinguishes nominal, ordinal, interval and ratio measurements?

Nominal values are unordered categories; ordinal values have order; interval values have meaningful equal differences; ratio values also have a meaningful zero.

Why should arbitrary category codes not silently enter a numerical distance?

Codes such as 1, 2 and 3 can create artificial ordering and spacing. Use an encoding or dissimilarity appropriate to the category meanings.

Does a dataset with no NA entries necessarily have no missing information?

No. Missing values may be stored as zero or another sentinel. Interpret the code using the variable meaning; zero can be valid for one variable and invalid for another.

What are the main tradeoffs of deleting or imputing missing values?

Deletion reduces data and may change the population represented. Mean/median imputation retains rows but compresses variation and does not restore missing uncertainty.

What is the z-score standardization formula?

z = (x - mean_train) / sd_train. Apply the training mean and standard deviation to validation/test observations; the standard deviation must be nonzero.

What is the min-max scaling formula?

x_scaled = (x - min_train) / (max_train - min_train). The training range must be nonzero. New values outside that range can scale below 0 or above 1.

Does z-score standardization make an arbitrary distribution normal?

No. It changes location and scale; it does not generally remove skewness, heavy tails or multimodality.

Why can scaling change KNN and K-means results?

Scaling changes feature contributions to distances. A variable with a large numerical scale can otherwise dominate neighbor selection or cluster assignments.

How do midpoint and median discretization differ?

The range midpoint is (minimum + maximum) / 2; the median is the middle position in the ordered data. Midpoint bins need not contain similar numbers of observations.

Do median cuts guarantee exactly 50% of observations in each category?

No. Ties and boundary conventions can prevent an exact 50/50 split. Balanced bins also do not guarantee useful outcome separation.

How many bins result from cuts at Q1 and Q3, and from cuts at Q1, median and Q3?

Q1 and Q3 give three bins. Adding the median gives four bins. Three cut points are not the same as three resulting intervals.

Why is one-hot encoding useful for an unordered categorical predictor?

It represents category membership through indicator columns without imposing an arbitrary numerical distance between category codes.

How do filter, wrapper and embedded feature selection differ?

Filters score features separately from the final fit; wrappers evaluate candidate subsets through model performance; embedded methods select while fitting, such as a coefficient penalty.

What is data leakage?

Information unavailable at the intended prediction time or outside the permitted training data influences fitting or selection, making evaluation too optimistic.

How can call duration leak information in a bank-marketing task?

If the decision is whom to call before a call starts, its eventual duration is unavailable. It can be useful for another analysis but leaks future information for that decision.

Where should scaling, imputation and supervised feature selection be fitted during cross-validation?

Inside each training fold. Apply the fitted transformations and selected features to that fold’s validation observations.

What conceptual R distinctions matter when reading a model script?

A factor represents categories; a data frame stores columns; NA marks missingness. Fitted values concern the fitted data, while predict with newdata evaluates supplied new observations.