If a dataset underrepresents certain groups, which bias is most likely to result?

• A. Representation bias
• B. Sampling bias
• C. Historical bias
• D. Proxy label bias

Answer: (A)

When some groups are not well represented in the data, the model learns mainly from the majority and those few examples, so its predictions end up mirrors of that dominant pattern rather than the true diversity of the population. This unequal representation biases the outcomes toward the well-represented groups and can reduce accuracy or fairness for the underrepresented ones. This situation is described as representational bias: the data’s composition misrepresents the population, shaping predictions in a biased way.

Sampling bias would involve the data collection process itself systematically favoring certain groups, which is related but focuses on how the data were gathered rather than the resulting composition of the dataset. Historical bias refers to prejudices or inequities embedded in the data from past real-world decisions, which can carry forward into models. Proxy label bias concerns using imperfect proxies for true labels, introducing bias in the labeling step rather than in the dataset’s representation.