Detectar y eliminar duplicados

Duplicate rows silently inflate counts, revenue totals, and averages without any error message. A sales dataset with 500 duplicate order records will overstate revenue by the exact sum of those 500 orders. Detecting and removing duplicates is one of the first cleaning steps you should perform, before any aggregation or modelling, because downstream errors compound from this single source of corruption.

df.duplicated() returns a boolean Series that is True for every row that is an exact copy of a previous row. The default keep='first' marks all but the first occurrence. Passing keep=False marks every occurrence of a duplicate — useful when you want to inspect all copies of a duplicated record before deciding which to keep.

Remove exact duplicate rows with df.drop_duplicates(). By default it keeps the first occurrence and drops all others. Pass keep='last' to keep the most recent record if your data has a timestamp and later rows are considered more authoritative. Always check the row count before and after to confirm the expected number of rows were removed.

Rows may be partial duplicates: they share a business key (like order_id) but differ in other columns due to an upstream bug that created two versions of the same record. Use df.duplicated(subset=['order_id']) to find rows with the same key but potentially different values in other columns. Inspect these rows before deciding how to reconcile them.

When partial duplicates exist, you must decide which record is authoritative. Common strategies: keep the row with the latest timestamp (most recent update), keep the row with the most non-null values, or keep the row with the higher dollar amount if one record has corrections. Sort by your tie-breaking criterion and then drop duplicates keeping the first (or last) occurrence.

Near-duplicates are rows that represent the same entity but differ slightly — for example, the same customer name with a typo, or the same transaction with a 1-cent rounding difference. Exact duplicate detection misses these. One approach is to group by the key column and count: if a customer name appears with slight variations, use .str.strip().str.lower() to normalise before deduplication.

When you need to merge duplicate rows rather than simply drop them, use groupby().agg(). For example, if two rows represent the same order but one has the shipping address and the other has the billing address, you can aggregate with 'first' (non-null preferred) or a custom function that coalesces non-null values across duplicates.

The result of drop_duplicates(keep='first') depends on row order. If the DataFrame was loaded from a database query without an ORDER BY clause, row order is non-deterministic, meaning the kept record may vary between runs. Always sort by a stable key (e.g. primary key or timestamp) before deduplication to make the process deterministic and reproducible.

After deduplication, verify the result with three checks: no remaining exact duplicates (df_clean.duplicated().sum() == 0), no duplicate business keys (df_clean.duplicated(subset=['order_id']).sum() == 0), and the expected row count is plausible. Write these as assertions so they fail loudly if future data changes invalidate the cleaning logic.

Good data cleaning is reproducible and auditable. Log the number of rows removed, the deduplication key, and the tie-breaking rule into a cleaning report dictionary. Persist this report alongside the clean data file so anyone running the pipeline later can verify the cleaning decisions without re-reading the code. Transparency in data cleaning builds trust in analysis results.

Save the deduplicated DataFrame to a new file rather than overwriting the original. This preserves the raw data for auditing. Name the file clearly with a _clean or _deduped suffix and include the run date so multiple cleaning passes are identifiable. Use Parquet for fast reloading in subsequent pipeline steps.

Test your understanding of Data Analysis concepts from this lesson.

In this lesson you learned: detecting exact and partial duplicates with duplicated() and drop_duplicates(), choosing which duplicate to keep using sorting and aggregation strategies, and verifying results with assertions and logging the cleaning decisions. Next up we explore outlier detection and treatment techniques.