← Lectures·L8

Data Bias: Modalities, Detection, and Mitigation

Prof. Dr. David B. Blumenthal · 6 concepts · 10 questions

Concept 1 / 6

What Is Data Bias? Definition, Consequences, and the Data-vs-Algorithmic Distinction

Data bias: a systematic distortion in the data. It arises from how data is collected, measured, labeled, or represented, and often reflects and perpetuates existing social, historical, or institutional inequities. It is typically unintentional and hard to detect without critical analysis.


Core principle: Systematic distortion in the data (data bias) can lead to systematic distortion in the results.


Bias vs. variance (archery-target analogy):

  • Bias = how far the cluster of shots sits from the bullseye (a systematic offset).
  • Variance = how scattered the shots are around their own center (random spread).

Real-world consequences:


DomainExample consequence
HealthcareDiagnostic algorithms misdiagnose certain populations due to non-representative training data
Criminal justicePredictive policing reinforces racial profiling and over-policing in marginalized communities
FinanceCredit scoring denies loans via proxies correlated with socio-economic background (e.g., zip code)
HiringResume screening perpetuates socio-economic, ethnic, or gender biases in historical hiring data

Bottom line: biased data is not just a technical issue — it is a societal and ethical challenge with tangible impacts on people's lives.


Data bias vs. algorithmic bias:


Data biasAlgorithmic bias
OriginThe data collection processHow algorithms process data
NatureReflects existing prejudices / systemic inequalitiesCan amplify or introduce new biases, even with unbiased data
ExamplesBiased survey responses, underrepresented groupsBiased recommenders, unfair credit-scoring models

In ML, where algorithms learn from data, data bias often leads to algorithmic bias.


Three modalities of data bias (the structure of this lecture): structural bias, measurement bias, and representational bias.