How is data prepared for machine learning?

How data is prepared for machine learning

Overview / motivation

• Real-world failure example: Amazon’s ML recruiting tool (reported by Reuters in 2018) was shut down after it learned to penalize resumes mentioning “women’s…” because the training data (10 years of resumes) was heavily male-dominated — a classic case of a faulty, biased dataset causing unfair outcomes.
• Key message: ML models learn from the data they’re given. Data quality, representativeness, and preparation determine success; bad data => bad models.

“Garbage in, garbage out”: models reflect the quality and biases of their training data.

---

High-level ML data-preparation workflow

1. Planning & problem formulation

   • Define the business problem to solve with ML before collecting and preparing data.

2. Data collection — size & examples

   • No one-size-fits-all for dataset size: collect as much relevant data as possible.
   • Examples:
     • Gmail smart-reply trained on ~238 million messages.
     • Google Translate used very large corpora (trillions of examples across languages).
     • Academic example: a Tamkang University professor used ~630 samples to predict concrete compressive strength.
   • Dataset size depends on task complexity and chosen algorithms.

3. Quality & adequacy of data

   • Accuracy depends on correctness and domain relevance of data.
   • Domain-mismatch example: using Canadian Thanksgiving sales to predict U.S. Thanksgiving turkey demand would be inadequate.

4. Labeling & features (supervised learning)

   • Labeling: provide “correct answers” so the model can learn (manual labeling or use existing labels).
   • Features: measurable characteristics describing examples (e.g., for an apple: shape, color, texture).
   • Common problem: mislabeled samples (e.g., peaches labeled as apples) — mitigate with cross-checking or multiple labelers.

5. Data reduction & cleansing

   • Dimensionality reduction: remove irrelevant, nearly-constant, or redundant features (e.g., drop country if all rows are US; drop year-of-birth if age is present).
   • Sampling: use subsets to speed prototyping and to rebalance classes.
   • Cleaning: fix or impute missing values (fill blanks with constants or predicted values) and remove corrupted or inaccurate records.

6. Data wrangling & normalization

   • Formatting: standardize file formats and categorical naming (e.g., “Florida” vs “FL”).
   • Normalization/scaling: unify numeric feature scales so features with larger numeric ranges don’t dominate (example: min–max normalization maps values to 0–1).

7. Feature engineering

   • Create new features from raw data to expose predictive signals (e.g., split datetime into day/month and hour to capture seasonality/time-of-day effects).
   • Thoughtful feature design can substantially improve model performance.

---

Practical notes & risks

• Class imbalance should be addressed (resampling, stratified sampling) to avoid biased models.
• Use cross-labeling and quality checks to reduce mislabeled data.
• Data preparation can consume up to ~80% of a data science project’s time.
• There are no flawless datasets; aim to reduce bias, improve representativeness, and clean/transform data to match the problem.

---

Resources referenced / examples

• Amazon’s ML recruiting tool (biased outcomes due to historical resume data) — Reuters report.
• Google examples:
  • Gmail smart-reply (~238 million messages)
  • Google Translate (very large corpora)
  • reCAPTCHA (used for label collection)
• Academic example: Tamkang University professor (used ~630 samples for a concrete strength prediction task)

Main speakers / sources

• Video narrator (overview and tutorial-style guide)
• Reuters (source for the Amazon recruiting tool story)
• Google (Gmail smart-reply, Translate, reCAPTCHA examples)
• Tamkang University professor (academic dataset example)

Note: the subtitles were auto-generated; some names/phrases may be slightly inaccurate.