mjhjmtu

I understand that when modeling, dummy variables should be k-1 and the dropped category should be the baseline. However, I do not know how to interpret if after feature selection 2 more categories of that dummy variable were removed (say I have a dummy variable with 5 categories - 1 would be the baseline, another 2 were removed after feature selection).

Should I still interpret it as usual, using the original dropped category as a baseline?

You should think of the k-1 dummy variables as a "block" - either they all stay in the model or they are all eliminated from the model during the feature selection process. The reason for this is that the k-1 dummy variables together help encode the effect of the original categorical variable that spawned them.

Suppose 5 categories are A, B, C, D and E. Suppose that A is for baseline, and C and E were removed in the process of variable selection.

It means A, C, and E have no statistically significant difference and they are combined into one group and treat A, C and E together as baseline.