Schema, provenance, missingness, duplicate and availability audits

1. Data inspection and correlation audit pattern

   Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2

   Book Chapter 2: Take a Quick Look at the Data Structure (printed pp. 54-57), Look for Correlations and Experiment with Attribute Combinations (pp. 65-69). Notebook 02: Take a Quick Look at the Data Structure (head, info, ocean_proximity value_counts, describe, hist) and Looking for Correlations (corr matrix sorted on median_house_value, scatter_matrix, median_income scatter) with Experimenting with Attribute Combinations (rooms_per_house, bedrooms_ratio, people_per_house)

   Done
2. Missing-value audit and the Explore checklist

   Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2

   Book Chapter 2: Clean the Data (printed pp. 70-73). Notebook 02: Data Cleaning: null-row audit with isnull().any(axis=1), the three options (dropna subset, drop column, median fillna), SimpleImputer on the numeric subset with statistics_ checked against manual medians; plus the checklist Explore the data steps (name, type, percent missing, noise, usefulness, distribution per attribute, target identification, documentation)

   Depth · Read the audit pattern here. The graded audit runs on Bike, not housing.

   Done
3. Bike dictionary and normalization provenance

   UCI Bike Sharing Dataset

   Variables Table roles, types, and descriptions; Additional Variable Information normalizations (temp and atemp min/max formulas on the hourly scale, hum divided by 100, windspeed divided by 67); Dataset Information no-missing-values statement; Dataset Files sizes (hour.csv 1.1MB, day.csv 56.2KB, Readme.txt)

   Done
4. Freshness and export checks as availability audits

   Rules of Machine Learning - first controls

   Rule 8: Know the freshness requirements of your system; Rule 9: Detect problems before exporting models

   Done

Focus on

• Audit in fixed order: shape and dtypes first (dteday loads as a string, not a datetime), then missingness (the page reports none and a direct count finds 0 nulls in 17379 hour rows), then keys and duplicates (instant unique, zero duplicated full rows), then time continuity (75 rows do not follow the previous clock hour, so the previous row and the previous hour differ), then ranges, normalizations, prevalence and availability.
• The checklist Explore list is the procedure: name, type, percent missing, noise, usefulness, and distribution per attribute, plus target identification and a sorted correlation pass. Document everything in the notebook.
• Unknowns and prevalence are decided before fitting. In the housing notebook, get_dummies builds whatever columns each frame holds, while the encoder fitted on the training set keeps its five categories; the test value <2H OCEAN, a misspelling of <1H OCEAN, is the unseen one. Bike weathersit takes all four values 1 to 4, but category 4 appears on only 3 of 17379 rows, so a split can easily miss it.
• A column present in the CSV may still be unavailable live. casual and registered are label parts, instant is a row id, and dteday must be parsed into cutoff-respecting parts. Record per-feature availability beside the audit.

Reproduce closed-book

Traps

• Trusting info and describe alone. describe skips strings (ocean_proximity needs value_counts) and hides the dteday-as-string problem; check dtypes explicitly.
• Reading normalized columns as raw units. temp 0.24 is not 0.24 C; invert \((t-t_{\min})/(t_{\max}-t_{\min})\) with the stated hourly minima and maxima (about 3.3 C for 0.24) before any physical claim.
• Counting the landing page as the audit. The page states 17389 instances with no missing values; verify both against the download (17379 rows, 0 nulls) and record version and date.

Practice

The section mini-project covers this: the full Bike audit suite (schema, types, provenance, missingness, duplicates, unknowns, prevalence, availability) with seeded defects that failing tests must catch.

Self-test

1. The imputed housing frame shows no nulls. How do you verify the imputer learned the right values?

   Show answer Got it Missed

2. temp reads 0.24 in hour.csv. What is it in Celsius?

   Show answer Got it Missed

3. get_dummies gives INLAND and NEAR BAY columns on one frame and <2H OCEAN and ISLAND columns on another. What breaks, and what does the notebook's fitted encoder do instead?

   Show answer Got it Missed

4. Which bike columns are unavailable as hourly-cnt features, and why?

   Show answer Got it Missed

Leads to

• 1.6 Training-only imputation, scaling, encoding and learned state Feed audited columns into train-only transformers with an unknown-category policy.
• 1.8 Snapshots, lineage, pipeline serialization and serving parity Snapshot the audited raw data and dictionary under versioned lineage.

Comes after

• 1.4 Leakage: target, future, entity, duplicate, preprocessing and feedbackBuild the audit suite that catches these defects routinely.
• 5.9 Covariate, label and concept shift, drift and label delayFeed drift monitors with audited schema and provenance.

Resources for this unit

• Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2
  Aurelien Geron (O'Reilly, 2025)primarycourse
  Read Chapters 1 and 2 of the book, then run notebooks 01 and 02 and use ml-project-checklist.md for framing, data inspection, test-set creation, preprocessing pipelines, cross-validation, error analysis and persistence. Rebuild the workflow on UCI Bike Sharing with a chronological split and prohibit casual and registered when predicting cnt.
• Rules of Machine Learning - first controls
  Google for Developerssupportapplication
  Read Rules 1-5 and 8-9. Turn the existing heuristic into a baseline and test the feature path independently of the learner.
• UCI Bike Sharing Dataset
  UCI Machine Learning Repositorysupportpractice
  Use hour.csv as the compact integrated dataset. Preserve a raw snapshot and data dictionary, split chronologically, and document why casual and registered are target proxies for cnt.