Log and power transformations: Question bank — AI-ML deep dive

Before we start, some plain-word reminders so every symbol below is earned:

True or false — justify

A power transform changes the ranking of your data points.

False. Log, sqrt, Box-Cox and Yeo-Johnson are all strictly increasing (monotonic) on their domains, so the largest value stays largest — only the spacing between points changes.

Applying

lo g

to a right-skewed feature always makes it perfectly Normal.

False. Log reduces right skew and often helps, but "reduce" is not "eliminate"; the result can still be skewed, bimodal, or over-corrected into left-skew. Normality is a goal, not a guarantee.

lo g (x + 1)

and

lo g (x)

give effectively the same result for large

x

True. When

x ≫ 1

, the

+ 1

is negligible:

lo g (1000001) \approx lo g (1000000)

. The offset only matters near zero, where it rescues

lo g (0)

from being undefined.

Box-Cox with

λ = 1

leaves the shape, spread, and ranks of the data unchanged.

True.

\frac{x ^{1} - 1}{1} = x - 1

is a pure shift by

- 1

— nothing about the distribution's form or ordering moves, so a fitted

λ \approx 1

signals "no transform needed."

Yeo-Johnson can always replace Box-Cox with no downside.

False. Yeo-Johnson handles zeros and negatives, but on strictly-positive data Box-Cox is often cleaner and more interpretable; Yeo-Johnson's piecewise form is harder to reason about, so use the simpler tool when it applies.

Log-transforming your target variable makes the model's raw predictions directly usable.

False. The model predicts in log space; you must inverse-transform with

exp (\cdot)

(plus a bias correction) to get back to original units.

A tree-based model like a decision tree benefits from log-transforming features.

False. Trees split on thresholds and only care about order; a monotonic transform gives identical splits, so it changes nothing for them. Transforms matter for models that assume a shape (linear/Gaussian), not for rank-based ones.

Square root is "gentler" than log for reducing right skew.

True. For

x > 1

lo g (x)

compresses large values more aggressively than

x

; sqrt preserves more of the original spacing, so it suits moderate skew.

The logit transform

lo g \frac{p}{1 - p}

works on any positive number.

False. Logit is for proportions strictly inside

(0, 1)

. At

p = 0

p = 1

the fraction blows up to

\pm \infty

(division by zero or

lo g 0

), so bounded proportions only.

Adding a constant before logging shifts the data but never distorts its shape.

False. The choice of constant changes the shape near small values:

lo g (x + 1)

lo g (x + 0.01)

pull the low end very differently. The constant is a modelling decision, not a free parameter.

Spot the error

transformed = np.log(prices) on a column that contains a 0.

Error:

lo g (0) = - \infty

, poisoning the column with -inf. Use np.log1p(prices) (which is

lo g (x + 1)

) so zeros map to

0

instead of blowing up.

Fit Box-Cox's

λ

on the entire dataset, then split into train/test.

Error: this is data leakage — the test set's distribution influenced

λ

. Split first, learn

λ

on train only, then apply that same

λ

to test.

"My target was log-transformed, so I'll report

exp (\overset{y}{^})

as the prediction."

Error:

exp

of the mean underestimates the true mean (Jensen's inequality). Add the bias correction

exp (\overset{y}{^} + \frac{σ ^{2}}{2})

, where

σ^{2}

is the residual variance in log space.

Run Box-Cox directly on temperature anomalies containing

- 8.2

Error: Box-Cox needs

x > 0

; a negative base with a fractional power is undefined. Use Yeo-Johnson, which mirrors the transform for

x < 0

"I'll scale with standardisation first, then log-transform."

Error: standardising creates negative values (mean-centred), breaking

lo g

. Transform first to fix the shape, then scale the already-symmetric result.

λ = - 1

gives

1/ x

, so it's the reciprocal transform."

Half-error: Box-Cox at

λ = - 1

\frac{x ^{- 1} - 1}{- 1} = 1 - \frac{1}{x}

, which is

- 1/ x

shifted — it preserves order (bigger

x

→ bigger output), whereas a bare

1/ x

reverses it. The

/ λ

and

- 1

matter.

Why questions

Why divide by

λ

and subtract

1

in Box-Cox instead of just using

x^{λ}

To make the family continuous at $λ = 0$ :

lim_{λ \to 0} \frac{x ^{λ} - 1}{λ} = lo g x

. Raw

x^{λ}

collapses to the constant

1

λ = 0

, leaving a hole exactly where log lives.

Why does log space turn multiplicative change into additive change?

Because

lo g (ab) = lo g a + lo g b

. A ×2 jump (

20 k \to 40 k

100 k \to 200 k

) is the same additive step

lo g 2

in log space, matching how we perceive proportional growth.

Why do we want skewness near zero after transforming?

A near-symmetric distribution better satisfies the Gaussian/linear assumptions of models like linear regression, and stops a few extreme values from dominating the fitted line.

Why does Yeo-Johnson use the exponent

2 - λ

(not

λ

) for negative inputs?

It's the mirror-image construction: it makes the negative branch match the positive branch's behaviour smoothly so the whole function is continuous and differentiable at $x = 0$ — no kink where the pieces meet.

Why prefer transformation over just deleting outliers?

Deletion throws away real information and can bias the model; a transform keeps every point but pulls extreme values inward, respecting that they may be genuine, meaningful observations.

Why is

lo g_{10}

sometimes preferred over natural log for interpretation?

Each unit in

lo g_{10}

means an exact ×10 change, which humans read easily ("two units = 100×"), whereas each unit in

ln

is a ×

e \approx 2.718

change that's harder to narrate.

Edge cases

What does

lo g (x + 1)

do to an input of exactly

x = 0

It maps to

lo g (1) = 0

— the clean, defined anchor point. This is precisely why log1p exists: to make zero a valid, well-behaved value.

Box-Cox at

λ = 0.5

on a value of

0

— problem or fine?

Fine in principle (

\frac{0 ^{0.5} - 1}{0.5} = - 2

), but Box-Cox's MLE for

λ

still requires strictly positive data, so a single zero forces you to add a constant or switch to Yeo-Johnson.

Yeo-Johnson at exactly

x = 0

with

λ \neq = 0

— which branch, and is it smooth?

The

x \geq 0

branch gives

\frac{( 0 + 1 ) ^{λ} - 1}{λ} = 0

; the transform equals

0

0

and the two pieces meet with matching slope, so it passes through the origin smoothly.

A feature is constant (every value identical) — what happens under any power transform?

The output is still constant (just a shifted constant); skewness is undefined and the feature carries zero information, so no transform can help — drop it instead.

What if the "optimal" fitted

λ

comes out very close to

1

The data is telling you it's already near-symmetric; applying Box-Cox does almost nothing, so skip the transform to keep the feature interpretable and avoid needless complexity.

Can a log transform ever increase skew?

Yes — on left-skewed data (long tail toward small values), log stretches the low end further and can worsen the asymmetry. Log is a right-skew remedy; diagnose the direction first.

Recall Fast self-test

Name the only transform in this note valid on negative inputs. ::: Yeo-Johnson. What must you do to log-transformed predictions before reporting them? ::: Inverse-transform with $exp$ , plus the $exp (\overset{y}{^} + σ^{2} /2)$ bias correction. Why do decision trees ignore monotonic transforms? ::: They split on order/thresholds, which monotonic maps leave unchanged.