Chapter 1 checkpoint

Ten questions about arrays, tables and trustworthy exploration.

Choose an answer before opening its feedback. The aim is to explain the reasoning, not memorise method names. If an answer surprises you, revisit the linked episode and rerun its example before continuing to Chapter 2.

Checkpoint

1. Units before arithmetic

You divide a mass in mg by a molar mass in g/mol, then by a volume in litres. What is the resulting concentration unit?

1. A mol/L

   Not quite. The mass was in milligrams rather than grams. Ignoring that factor would make the numerical concentration wrong by 1,000.

2. B mmol/L

   Correct. Correct: mg divided by g/mol gives mmol, then dividing by litres gives mmol/L. The function in episode 1.1 documents this convention.

3. C g/L

   Not quite. Dividing by molar mass converts a mass amount into a molar amount; the numerator is no longer a mass.

Checkpoint

2. Which axis?

X.shape is (120, 5): 120 molecules and five descriptors. What does X.mean(axis=0) return?

1. A Five means, one per descriptor

   Correct. Correct: axis 0 is reduced, leaving the five columns. The shape is (5,), which can broadcast against (120, 5) to centre columns. See episode 1.2.

2. B 120 means, one per molecule

   Not quite. That is X.mean(axis=1), which reduces the feature axis. Averaging unlike descriptor units may not have a scientific interpretation.

3. C A single mean of all 600 values

   Not quite. That is what X.mean() does without an axis; it also mixes unlike descriptor units.

Checkpoint

3. A slice changes the original

After part = X[:10], assigning part[0, 0] = 0 also changes X. Why?

1. A NumPy has lost track of the variable names

   Not quite. Names are working normally. Two arrays can refer to shared underlying data.

2. B A basic slice is a view of the same data

   Correct. Correct. Use part = X[:10].copy() when you want independent data. This rule concerns basic slicing; boolean and integer-array selection have different copying behaviour.

3. C Every NumPy selection always makes a copy

   Not quite. If this were a copy, writing to it would not change X. Basic slices normally share data for efficiency.

Checkpoint

4. Labels and positions

A filtered DataFrame has row labels [10, 25, 40]. Which expression selects its second row by position?

1. A df.loc[1]

   Not quite. loc asks for the row labelled 1, which is absent. It does not count from the top.

2. B df.iloc[1]

   Correct. Correct: positions start at zero, so position 1 is the second row, labelled 25 here. Episode 1.3 distinguishes labels from positions.

3. C df.iloc[25]

   Not quite. This asks for the 26th row by position, outside a three-row table. df.loc[25] would select the desired label.

Checkpoint

5. A merge with unchanged row count

A left join returns the same number of rows as its left input. Does that prove every input row matched the lookup table?

1. A Yes, every unmatched row would have disappeared

   Not quite. A left join retains unmatched left rows and fills right-hand columns with missing values. An inner join would discard them.

2. B No; inspect the match indicator and validate the lookup key

   Correct. Correct: indicator=True records matches, and validate="many_to_one" checks that the right-side key is unique. An unchanged row count alone does not prove complete matching.

3. C No; use an inner join to hide the unmatched rows

   Not quite. Hiding unmatched rows changes the analysed sample without explaining the problem. Investigate the keys and document any exclusions.

Checkpoint

6. Missing is not zero

Five percent of assay readings are absent, especially among poorly soluble compounds. What is the best first action?

1. A Replace every missing reading with zero

   Not quite. Zero is a measurement with a meaning. Using it as a placeholder changes the distribution and can create a false relationship.

2. B Delete the rows because five percent is small

   Not quite. A small loss can still be selective. Here missingness is related to a scientific property, so deletion may change the population being described.

3. C Investigate and document the missingness before choosing a rule

   Correct. Correct: examine available metadata and the intended analysis. If an imputation method is later used for prediction, fit it only on the relevant training data.

Checkpoint

7. What do error bars mean?

A plot labels its error bars as standard deviation across independently sampled birds. What do they describe?

1. A The spread of individual bird measurements

   Correct. Correct: standard deviation describes variability in measurements. It is different from uncertainty in an estimated mean; episode 1.4 explains why error bars need explicit labels.

2. B A 95% confidence interval for the population mean

   Not quite. A standard deviation is not automatically a confidence interval. Coverage requires an interval construction and assumptions about the sampling process.

3. C The error rate of a classifier

   Not quite. No predictions are being evaluated here. These bars summarise the observations themselves.

Checkpoint

8. Reading a logarithmic target

Two compounds have logS values of −3 and −5. How do their recorded molar solubilities compare?

1. A The first is two times as soluble

   Not quite. A difference of two on a base-10 logarithmic scale means two powers of ten, not a factor of two.

2. B The first is 100 times as soluble

   Correct. Correct: the ratio is 10**(-3) / 10**(-5) = 100. Episode 1.5 interprets differences in logS as ratios of solubility.

3. C The second is 100 times as soluble

   Not quite. The more negative logS denotes the smaller positive concentration, so the direction is reversed.

Checkpoint

9. Exploration and evaluation

You rank features using the labels of every row, keep the top five, and only then reserve 20% as a test set. What is the problem?

1. A There is no problem because the model has not been fitted yet

   Not quite. Feature selection is already a modelling decision. The future test labels influenced it before the model fit.

2. B Test information has already influenced the selected features

   Correct. Correct: reserve test data before target-guided decisions. During cross-validation, fit feature selection inside each training fold. The whole-table EDA in Chapter 1 teaches exploration; it does not create independent confirmation of a model designed from those plots.

3. C The problem is solved by choosing seed 42

   Not quite. A fixed seed makes the split repeatable; it does not remove information that has already influenced a decision.

Checkpoint

10. From pattern to explanation

Molecular weight and logS are negatively correlated in this dataset. Which conclusion is supported by that observation alone?

1. A Adding any atom will necessarily lower solubility

   Not quite. That is a causal claim about a structural intervention. The correlation does not control what atom, functional group or other property changes.

2. B Every heavier molecule is less soluble than every lighter molecule

   Not quite. A population trend does not determine every pair of observations. The scatter plot has substantial overlap.

3. C Higher molecular weight is associated with lower logS in this sample

   Correct. Correct: this states the measured association and its scope. Predictive usefulness must be evaluated on suitable held-out data, and a causal explanation requires more than correlation.

Ready to move on

You are ready when you can explain a row’s identity and units, inspect a table without silently changing its meaning, and distinguish a pattern worth investigating from an independently tested conclusion. Lab 1 puts these habits into practice; Chapter 2 turns them into an evaluation workflow.

PreviousLab 1 Guided EDA: the breast cancer Wisconsin datasetNextChapter 1 key points

Edit this pageReport a problemLast updated 2026-11-04