Descriptive Statistics for ML: Mean, Variance, IQR and Robust Summaries

Work through one seven-value feature to see how ordinary and robust summaries react to a long right tail. Then turn the results into careful preprocessing choices.

KnowledgeGate Team

Exam prep & CS education

Updated 11 Nov 20266 min read
Descriptive Statistics for ML: Mean, Variance, IQR and Robust Summaries

A single average can make a feature look ordinary even when most observations cluster tightly and one value stretches the scale. Consider delivery delays [2, 3, 3, 4, 5, 7, 30] minutes: six are within 7 minutes, but one is 30. Computing centre, ordinary spread and robust spread by hand helps decide what to inspect before an ML transformation or model. The 30-minute observation stretches the mean and standard deviation while leaving the median unchanged, so mean and standard deviation are insufficient here.

Descriptive statistics are a feature audit, not one magic number

Descriptive statistics compress observed data. Centre covers mean, median and mode; spread covers range, variance and standard deviation; robust summaries include median, quartiles, IQR and median absolute deviation (MAD). Compression hides detail, so also report units, count and shape. The Skills category offers broader applied technical learning.

Observation

Delay (minutes)

D1

2

D2

3

D3

3

D4

4

D5

5

D6

7

D7

30

Here, n = 7, minimum 2, maximum 30, and six values lie from 2 to 7. The value 30 is unusual here, but summaries cannot blame data, traffic, weather or another delivery class.

Mean, median and mode answer different centre questions

Mean numerator:

2 + 3 + 3 + 4 + 5 + 7 + 30 = 54

Thus mean = 54 / 7 = 7.714285..., or 7.71 minutes. It is the balance-point delay, not necessarily a typical delivery's experience.

The sorted list's median position is (7 + 1) / 2 = 4, so median = 4 minutes. Mode is 3 minutes because it appears twice. Here it is just a repeated observed value, not a measure that always suits continuous data.

Mean, median and mode are 7.71, 4 and 3 minutes. The 30-minute value lifts the mean without changing the middle position. With the visible long right tail, mean > median is consistent with right skew here, but the inequality alone never proves skew.

Range, variance and standard deviation quantify ordinary spread

Range is 30 - 2 = 28 minutes. It ignores the five values between its endpoints and remains measured in minutes.

For this batch, use population variance with mean 54/7 and denominator n = 7:

Value

Squared deviation from 54/7

2

32.6531

3

22.2245

3

22.2245

4

13.7959

5

7.3673

7

0.5102

30

496.6531

They sum to 595.4286. Population variance is 595.4286 / 7 = 85.0612 minutes squared; population standard deviation is sqrt(85.0612) = 9.2229 minutes. Only SD returns to the original unit.

If these records are a sample estimating a wider population, divide by n - 1 = 6: sample variance 99.2381 minutes squared, sample SD 9.9618 minutes. State the denominator and convention.

Median, IQR and MAD stay stable when the tail stretches

Using Tukey's median-of-halves convention, exclude the overall median because n is odd. Lower half [2, 3, 3] gives Q1 = 3; upper half [5, 7, 30] gives Q3 = 7; IQR = 7 - 3 = 4 minutes. Software interpolation can differ on small datasets, so record the method.

The 1.5 x IQR fences are 3 - 1.5(4) = -3 and 7 + 1.5(4) = 13. Since 30 > 13, investigate it. A fence is diagnostic, not proof of error or a deletion command.

Absolute deviations from median 4 are [2, 1, 1, 0, 1, 3, 26]; sorted, [0, 1, 1, 1, 2, 3, 26]. Their median gives raw MAD = 1 minute. Some workflows rescale MAD, so label any scaled convention.

Dot plot and corrected boxplot with right whisker ending at Q3 7 and 30 marked beyond the upper fence 13.

One sensitivity check reveals which summaries are robust

Change only the tail value from 30 to 60:

Summary

Original tail 30

Changed tail 60

Mean

7.71

(54 - 30 + 60) / 7 = 12

Median

4

4

Range

28

60 - 2 = 58

IQR

4

4

Raw MAD

1

1

Mean rises by 4.29 minutes and range by 30 minutes; median, IQR and raw MAD stay fixed. Robustness means limited sensitivity to extremes, not superiority. For symmetric, clean data, mean and SD retain every observation and may fit better.

Report: n = 7; median = 4 minutes; IQR = 4 minutes; mean = 7.71 minutes; SD = 9.22 minutes; one value, 30, lies above the 13-minute Tukey fence. Seven records cannot establish a stable population shape or choose a model.

Turn the summaries into a defensible ML preprocessing decision

Investigate before deleting. Check the record's units, collection logic, missing-value encoding, duplicate status and context. Correct a genuine entry error with an audit trail; retain a valid rare event. Never remove it just to reduce the mean.

Standard scaling with mean 7.7143 and population SD 9.2229 maps 30 to (30 - 7.7143) / 9.2229 = 2.42. Robust scaling with median 4 and IQR 4 maps values to [-0.50, -0.25, -0.25, 0, 0.25, 0.75, 6.50]. It stops the tail defining centre and scale, but neither erases it nor guarantees a better model.

One candidate, log1p(x), maps these non-negative values to [1.0986, 1.3863, 1.3863, 1.6094, 1.7918, 2.0794, 3.4340], with mean 1.8265 and median 1.6094. It changes interpretation, so justify it through feature meaning and validation. Fit transformations on training data, then apply them unchanged to validation and test data.

Dependency-free Python keeps the conventions explicit:

python
from math import sqrt

x = sorted([2, 3, 3, 4, 5, 7, 30])
n = len(x)
mean = sum(x) / n
variance = sum((value - mean) ** 2 for value in x) / n
sd = sqrt(variance)
median = x[n // 2]
lower = x[:n // 2]
upper = x[n // 2 + 1:]
q1 = lower[len(lower) // 2]
q3 = upper[len(upper) // 2]
iqr = q3 - q1
absolute_deviations = sorted(abs(value - median) for value in x)
mad = absolute_deviations[n // 2]

print(f"mean {mean:.4f}, variance {variance:.4f}, sd {sd:.4f}")
print(f"median {median:.4f}, Q1 {q1:.4f}, Q3 {q3:.4f}")
print(f"IQR {iqr:.4f}, MAD {mad:.4f}")

Output is mean 7.7143, variance 85.0612, sd 9.2229; median 4.0000, Q1 3.0000, Q3 7.0000; IQR 4.0000, MAD 1.0000. The Python course builds the programming foundation this snippet assumes.

Raw delivery delays aligned with their log1p values, showing how the transformation compresses the 30-minute right tail.

Assessment and interview-style questions test calculation plus judgement

Stable question forms include computing a missing summary, finding the measure most changed by an extreme, distinguishing variance from SD units, choosing ordinary or robust summaries, detecting denominator mismatch, and explaining why a fence does not prove error.

Four rapid checks:

  1. If 30 becomes 60, what happens? Mean becomes 12; median remains 4.

  2. Is 30 flagged by Tukey's rule? Yes. The upper fence is 13.

  3. What are the two variances? Population 85.0612; sample 99.2381.

  4. What is robust-scaled 30? (30 - 4) / 4 = 6.50.

Trap

What goes wrong

Repair

Call 7.71 typical

Hides median 4

Report both

Write SD as 85.06

Confuses variance and SD

Use SD 9.22

Compare variance with raw data

Units differ

Label squared minutes

Use n - 1 for population variance

Denominator and label conflict

Use n, or relabel sample variance

Accept a software quartile silently

Convention is unknown

Record Tukey's method

Delete 30 because it crosses 13

A flag becomes a verdict

Investigate context

Scale before splitting

Validation information leaks

Fit on training only

Probability for GATE CS moves from describing observations to reasoning about uncertainty.

Descriptive statistics for ML: the short version and next step

  • Inspect units and raw values.

  • Report n.

  • Compute more than one centre.

  • Pair range with variance or SD.

  • Add median, IQR and MAD when tails matter.

  • Investigate unusual records before transforming.

Headline: mean 7.71, median 4, population SD 9.22, IQR 4, raw MAD 1. Without looking back, reproduce sum 54, fourth-position median 4, upper fence 13 and raw MAD 1, then explain why 30 changes mean more than median.

For the wider domain, work through the Artificial Intelligence (AI) course. For another mathematical pillar of ML, study Linear Algebra for GATE CS. Next, run this template on a real feature, record the quartile convention, and validate any transformation.