Machine Learning Basics
When building a model that predicts hiring decisions from application documents, using past hiring outcomes as the correct answers, which of the following should be kept in mind?
- AIf past decisions were biased by gender or other factors, the model may learn that bias and reproduce or amplify it
- BIf there is enough data, the biases in the original data cancel out, so the model's predictions naturally become fair
- CIf the gender column is removed from the input beforehand, bias cannot enter the model through other features
Show answer and explanation
Answer: A. If past decisions were biased by gender or other factors, the model may learn that bias and reproduce or amplify it
A machine learning model learns the tendencies in its training data as they are, so it also learns biases in the data, such as discrimination contained in past decisions or a shortage of data for particular groups. Adding more biased data does not remove the bias itself, and even with the gender column removed, bias can remain through features closely related to it, such as career history or hobbies. It is important to review how the data was collected, check error rates for each group and evaluate fairness.
- A. Correct. The model can learn the bias in the data and reproduce it as is.
- B. However much biased data is added, the bias does not disappear.
- C. Bias can remain through other features closely related to gender (proxy variables).