Types of Bias in Machine Learning
What you'll learn: You'll understand the four primary types of bias that can corrupt ML systems—selection, measurement, label, and aggregation bias—and how each distorts outcomes.
Understanding the Four Core Biases
Machine learning systems don't create bias from nowhere—they inherit and amplify biases introduced at different stages of development. Let's examine each type:
Selection Bias
This occurs when your training data doesn't represent the real-world population your model will serve. Analogy: Training a face recognition system exclusively on well-lit professional photos, then deploying it in varied lighting conditions and angles. The system fails because it never "saw" realistic scenarios during training.
Impact: Models perform poorly on underrepresented groups, perpetuating inequality.
Measurement Bias
Happens when your data collection instruments or processes systematically distort reality. For example, using smartphone sensors to detect health conditions disadvantages users without premium devices, or arrest records as proxies for crime rates (which reflect policing patterns, not actual crime).
Impact: The model learns correlations that reflect measurement flaws rather than truth.
Label Bias
Emerges when human labelers inject their prejudices into training data annotations. If resume screeners consistently rate identical resumes differently based on perceived gender or ethnicity, the model learns discriminatory patterns.
Impact: The model replicates and automates human prejudice at scale.
Aggregation Bias
Occurs when you treat distinct subgroups as a single population, assuming one model fits all. A diabetes risk model trained on aggregated data might work well on average but fail dramatically for specific ethnic groups with different physiological patterns.
Impact: Vulnerable populations receive inaccurate predictions while aggregate metrics look acceptable.
Why This Matters
These biases often compound—selection bias in who gets loans creates biased training data for future loan models, creating a self-reinforcing cycle of exclusion.
Key Takeaway: The four types of ML bias—selection, measurement, label, and aggregation—enter systems at different stages but share a common danger: amplifying historical inequities through automated decision-making at scale.