Machine learning models learn from examples. If those examples are incomplete, outdated or unrepresentative, the model learns the wrong lessons, no matter how advanced its design.

Common data problems

  • Missing or incorrectly recorded values.
  • Duplicate or inconsistent records.
  • Labels applied carelessly or differently by different people.
  • Data that covers only some groups, places or situations.
  • Information that is out of date for the problem being solved.

How poor data shows up

A model trained on skewed data may perform well in testing yet fail in the real world. It can also treat some groups unfairly if they were under-represented or described inaccurately in the training set.

Good practices

Teams usually start by exploring and cleaning data, documenting where it came from and how it was labelled. They split data into training and testing sets, check performance across different groups, and keep monitoring the model after launch, since real-world patterns change.

Privacy and consent

Collecting data responsibly matters too. Organisations should gather only what they need, protect it properly and follow relevant privacy laws.

Takeaway

Before asking how clever a model is, ask what data it learned from. Careful, representative and well-documented data is the foundation of trustworthy machine learning.