What is representation bias?

What is representation bias?

• A. Bias caused when some groups are underrepresented in the data used to build/evaluate a model
• B. Bias caused when some groups are overrepresented
• C. Bias introduced by sampling errors
• D. Bias due to measurement errors

Answer: (A)

Representation bias arises when the data used to train or evaluate a model do not reflect the full diversity of the real world, so some groups are underrepresented. This gives the model less information about those groups and leads to biased predictions or outcomes for them. It’s a systematic issue rooted in how data are collected and sampled, not random error. If a model is trained mostly on one demographic, it may perform well for that group but poorly for others. Overrepresentation would skew results in the opposite direction, but representation bias specifically refers to underrepresentation. Sampling errors are random fluctuations from taking a subset, not a biased undercount of groups, and measurement errors are about inaccuracies in the data values themselves, not how many examples come from each group.

Representation bias arises when the data used to train or evaluate a model do not reflect the full diversity of the real world, so some groups are underrepresented. This gives the model less information about those groups and leads to biased predictions or outcomes for them. It’s a systematic issue rooted in how data are collected and sampled, not random error.

If a model is trained mostly on one demographic, it may perform well for that group but poorly for others. Overrepresentation would skew results in the opposite direction, but representation bias specifically refers to underrepresentation. Sampling errors are random fluctuations from taking a subset, not a biased undercount of groups, and measurement errors are about inaccuracies in the data values themselves, not how many examples come from each group.