Follow the transformation
Measure the problem by column, row group and time/segment.
Investigate the data-generating process before choosing a fix.
Choose deletion, correction, imputation, transformation or retention with an explicit reason.
EDA & Data Quality · Lesson 25
Outliers
Outliers is a data-quality decision, not merely a cleaning command.
What Outliers actually means
Outliers is a data-quality decision, not merely a cleaning command. The correct treatment depends on how the issue arose, whether it carries information, and how the treatment changes the population or downstream model.
Outliers matters because model quality cannot exceed the meaning and integrity of its data. Profiling, EDA, label checks and leakage checks expose problems that a sophisticated algorithm may otherwise exploit or hide.
Deeper walkthrough
Read Outliers as a mechanism, not a recipe
Treat this as a sequence of observable decisions rather than one opaque command. Stage 1: Measure the problem by column, row group and time/segment. Stage 2: Investigate the data-generating process before choosing a fix. Stage 3: Choose deletion, correction, imputation, transformation or retention with an explicit reason. Final checkpoint: Record an indicator or audit trail when the fact that a value was missing/changed may itself matter.
Know what would convince you
- Compare missingness, distributions, row counts and key constraints before and after the change.
- Inspect representative changed rows and confirm the rule with domain/data documentation.
Useful distinctionDeletion: Removes affected rows/columns; can bias the population.
Click a stage to inspect what happens, what changes, and what should be checked before moving on.
Stage 1
Measure the problem by column
Measure the problem by column, row group and time/segment. For Outliers, identify the exact state before this stage, the operation or rule applied here, and the observable state afterwards so the mechanism remains inspectable.
State focus: identify exactly what changed at this stage and what observable evidence confirms that change.
Trace the mechanism step by step
- Measure the problem by column, row group and time/segment.
- Investigate the data-generating process before choosing a fix.
- Choose deletion, correction, imputation, transformation or retention with an explicit reason.
- Fit learned preprocessing only on training data when predictive modelling is involved.
- Record an indicator or audit trail when the fact that a value was missing/changed may itself matter.
Make the concept concrete
Demonstration
Python / NumPy example
# Step 1 — Import the module so its functions/classes are available to the rest of this example.
import numpy as np
# Step 2 — Construct `x` as an array so vectorised numerical operations can be applied consistently.
x = np.array([10,11,11,12,12,13,40])
# Step 3 — Compute the right-hand expression and store its result in `q1,q3` for the next step.
q1,q3 = np.quantile(x,[.25,.75]); iqr=q3-q1
# Step 4 — Compute the right-hand expression and store its result in `lo,hi` for the next step.
lo,hi=q1-1.5*iqr,q3+1.5*iqr
# Step 5 — Display the current value explicitly so the result/state can be inspected during execution.
print("fences:",lo,hi)
# Step 6 — Display the current value explicitly so the result/state can be inspected during execution.
print("flagged:",x[(x<lo)|(x>hi)])Expected / illustrative result
40 is flagged by the 1.5×IQR rule. That is a review flag, not automatic evidence the value is wrong.
Interpret the result.
For Outliers, trace representative source rows/columns into the result and reconcile row counts, dtypes, keys or missing values that the operation could change.
Know what this is — and what it is not
DeletionRemoves affected rows/columns; can bias the population.Simple imputationUses a fixed statistic/category; easy but shrinks variability.Model-based imputationUses relationships with other variables; more assumptions and leakage risk.Missing indicatorPreserves information about whether a value was missing.When it is appropriate
Use Outliers when it helps diagnose, document or correct a data-quality issue without destroying information needed for the downstream question.
When to stop or reconsider
Do not “clean” automatically when the apparent anomaly may carry signal, reflect data collection, or require domain adjudication; preserve an audit trail of changes.
Failure modes to recognise
- Treating missing values, outliers or labels as purely technical defects without investigating how they were generated.
- Applying learned cleaning/imputation using information from validation/test data.
- Changing values without recording which rows changed and how distributions/counts were affected.
How to check the result
- Compare missingness, distributions, row counts and key constraints before and after the change.
- Inspect representative changed rows and confirm the rule with domain/data documentation.
- In predictive work, fit learned cleaning only on training data and verify the pipeline reproduces that boundary.
Demonstrate understanding
Try this:
Build a tiny, inspectable example of Outliers. First measure the problem by column, row group and time/segment. Then investigate the data-generating process before choosing a fix. Write the expected result before running it, and explain one condition that would make the result misleading or invalid.
Start by measuring the problem, not fixing it. Keep a before/after table of counts or distributions and inspect the exact records affected by the rule.
Check reasoning, not memorisation
Before trusting a result from Outliers, which check provides the strongest evidence that you understand and applied it correctly?
Keep the important distinctions visible
Step 1Measure the problem by column, row group and time/segment.Step 2Investigate the data-generating process before choosing a fix.Step 3Choose deletion, correction, imputation, transformation or retention with an explicit reason.Step 4Fit learned preprocessing only on training data when predictive modelling is involved.What to remember
- Outliers is a data-quality decision, not merely a cleaning command. The correct treatment depends on how the issue arose, whether it carries information, and how the treatment changes the population or downstream model.
- Measure the problem by column, row group and time/segment.
- Treating missing values, outliers or labels as purely technical defects without investigating how they were generated.
- Compare missingness, distributions, row counts and key constraints before and after the change.