I am experimenting with Random Forests for a Classification problem and am wondering about calibration for individual features. For example, a 2-level factor such as home field advantage in a sporting event, that we can be certain has an average effect of around +5% on win rate and its effect is not captured by any other feature in the data.
It appears as though the nature of random forests (selecting N random features to consider at each split) will not allow the model to fully capture the effect of any one particular feature like this. Setting the max_features parameter to None or to include all features appears to solve the problem but then it loses the power of diversity between trees.
I am wondering if there are any good methods of dealing with this type of feature that we want to be fully captured based on some sort of domain knowledge we have about the problem?