Metadata describes what data fields mean and how they were produced. A useful description might identify a column’s units, expected decimal places, and the script that generated it. Without that context, a column called “sales” could mean currency or physical quantity. Read the available metadata before calculating anything, then inspect the file itself. section 2.3
Do not let a plausible column name become an assumption. An “address” might belong to a customer or a shop; a “time” might record an order, a purchase, or a test value. When a definition matters to a business decision, ask the people responsible for the data. Misdefined message-opening and receipt events once produced impossible ratios and invalidated earlier efficiency analyses. section 2.3
When documentation is incomplete:
- Inspect the raw file and available metadata before interpreting its columns. section 2.3
- Form a testable interpretation of an unclear field and compare it with a related field. For two review timestamps, calculate their differences and inspect the resulting distribution. section 2.6.3
- Treat the result as evidence, not confirmation. Two time-based columns may correlate simply because their records follow the same timeline; seek a reliable definition before using an uncertain field in analysis. section 2.6.3
df_order_reviews['time_diff'] = (df_order_reviews['review_answer_timestamp'] – df_order_reviews['review_creation_date']).dt.total_seconds()
Subtracting review creation time from answer time creates a value that can be summarized and plotted to test a possible interpretation of the timestamps. It does not, by itself, establish their official meanings. section 2.6.3
