pandas-data-cleaner

Overview

Provides a complete, reusable pandas data cleaning pipeline for raw datasets. Covers missing value strategies (drop, fill, interpolate), duplicate detection and removal, dtype inference and safe casting, string normalization, outlier detection (IQR, Z-score, isolation forest), column renaming conventions, and a full end-to-end cleaning function template that can be dropped into notebooks or scripts.

When to Use This Skill

• Receiving raw data from exports, APIs, or databases that needs preparation.
• Exploratory data analysis (EDA) before modeling.
• Building reproducible data preparation steps for ML or reporting.

Prerequisites

• Python environment with pandas, numpy, and optionally scikit-learn for advanced outlier detection.
• The raw dataset (CSV, Parquet, DataFrame in memory, or SQL query result).
• Understanding of the domain (what "valid" looks like for each column).

Steps

1. Initial inspection:
   • df.head(), df.info(), df.describe(include='all'), df.isnull().sum().
   • Identify obvious issues (mixed types, weird strings, extreme values).
2. Missing values strategy (per column or group):
   • Drop rows/columns if <5% missing and not critical.
   • Fill with median/mean/mode for numeric.
   • Forward/backward fill or interpolate for time series.
   • "Unknown" or sentinel value for categorical.
   • Flag rows with missing critical fields.
3. Duplicates:
   • Exact duplicates: drop_duplicates().
   • Fuzzy duplicates (same entity different spelling): use recordlinkage or simple key normalization + group.
4. Dtype cleaning & casting:
   • Infer dtypes safely.
   • Convert dates with pd.to_datetime(errors='coerce').
   • Numeric with pd.to_numeric(errors='coerce').
   • Categorical for low-cardinality strings.
5. String normalization:
   • .str.strip().str.lower().
   • Remove accents, normalize unicode.
   • Standardize categories (map "USA", "U.S.", "United States" → "US").
6. Outlier detection & treatment:
   • IQR method for most numeric.
   • Z-score for near-normal distributions.
   • Domain-specific rules (negative age, price > 10x median, etc.).
   • Cap, clip, or flag rather than blindly remove.
7. Column hygiene:
   • Rename to snake_case, consistent naming.
   • Reorder columns logically.
   • Add derived columns (e.g., is_weekend, age_group).
8. Validation:
   • Assert expected columns, dtypes, row count range, no nulls in critical fields.
   • Great Expectations or custom checks.
9. Output:
   • Reusable clean_data(df: pd.DataFrame) -> pd.DataFrame function.
   • Before/after summary statistics.
   • Notebook cells or script that can be run on new files.
   • Export to Parquet/CSV with metadata.

Examples

A complete cleaning pipeline for a typical customer/orders dataset (handle missing emails, normalize country codes, detect price outliers, deduplicate on customer_id + email, cast dates, add features) with before/after reports is included.

Edge Cases & Error Handling

• Mixed types in a column: Coerce and report how many rows became NaN.
• Very large files: Use chunksize or Dask / Polars for out-of-core.
• Time series specific: Resample, forward fill with limits, detect gaps.

Verification

1. Run the cleaner on the raw data — no errors.
2. clean_df.isnull().sum() shows only expected missing values.
3. clean_df.dtypes are correct.
4. Spot checks: random samples look reasonable.
5. Row count and key statistics are within expected ranges (or documented changes are explained).
6. Re-run on a new batch of raw data produces consistent output shape.
7. Success: The cleaned dataset is ready for analysis or modeling with documented transformations.

References

• Pandas User Guide - Working with missing data
• Pandas Best Practices
• Data Cleaning with Python (book)