Cleaning steps alone are not enough. The important question is what changed, why it changed, and whether the cleaned dataset is safer to analyze.
1. Profile before cleaning
Record shape, data types, missingness, duplicate counts, unique values, and basic distributions. This gives you a baseline.
2. Understand field meaning
A null can mean unknown, not applicable, not collected, or a pipeline failure. Do not choose a fill strategy without understanding the field.
3. Standardize types
Convert dates, numeric values, booleans, and categories deliberately. Type errors can silently break filters and aggregations.
4. Investigate duplicates
Identify the business key and determine whether repeated rows are legitimate events, updates, or accidental duplication.
5. Normalize categories carefully
Create explicit mappings for important category standardization. Avoid blind replacements that can merge distinct values.
6. Treat outliers as signals
An extreme value may be real, a unit mismatch, an entry error, or a system issue. Investigate before deleting.
7. Validate after transformation
Compare row counts, key counts, totals, distributions, and missingness before and after cleaning. Document material changes.
Ready-for-analysis checklist
The dataset should have a known grain, documented assumptions, validated transformations, understood missingness, and a clear record of what was changed.