I was thinking today about how biased data affects machine learning performance, and I begin to wonder why class imbalance (or data imbalance in general) is only talked about in a classification output variable context.
Why is imbalanced data not really talked about in an input variable context? If my sample data used for training has 70% of A and 30% of B for a given input variable, and we happen to skew A to correlate with X in the training data, is this not a problem?
Would this also be an explanation why machine learning algorithms discriminate against certain groups of people? Like in criminal justice machine learning applications where a criminal sample is used to represent the entire population, and a certain group is overrepresented in the sample?
If this concern is true, then the ideal machine learning dataset is just a combinatorial of every possible input variable with the desired output variable right?