In this project I compare how model explainability changes if data are biased. I consider 2 datasets (biased and unbiased) - the biased dataset is created by filtering the unbiased.
Similar accuracy can suggest that models work similarly but in features importance analysis (LIME, Permutation Importance, Feature Importance for Random Forest, Partial plot importance) we can see that biased one use feature sex as one of most important variables.
Dataset: http://archive.ics.uci.edu/ml/datasets/default+of+credit+card+clients
For detailed analysis please check my blog: https://analitycznyumysl.pl/kiedy-ufac-modelom-ml/