Generally model gets biased towards data_samples/target whose frequency is high in training data set. Is it possible during training that model gets biased towards low frequency training data set.
With structured data, you have in general 4 challenges:
(1) Missing data
(4) Rare values (as a rule of thumb <5%)
Rare values in categorical variables tend to cause over-fitting, particularly in tree based methods. Ph.D. Data Scientist Soledad Galli has an amazing course on the subject (Udemy: "Feature Engineering". Below a screenshot from her course, but to be fair to her, I'm not going to post the solution.