I have a very strange problem when trying to use random forests. I am trying to use RF for some binary image segmentation where I am using some texture features and I am using the scikit-learn library for this.
So, since I have a bit of unbalanced classes issue (most of the image is background), I use the
balanced_subsample parameter and fit the model as:
model = RandomForestClassifier(n_estimators=250, max_depth=12, random_state=42, verbose=3, class_weight='balanced_subsample')
Th thing I am seeing is that when I segment an image using this fitted model, the segmentation would be rubbish but if I just flipped the predictions, it would be good. It is almost as if the model has learnt the inverted labeling.
Could this be something that can happen with random forests? I have checked my code and verified that the labels are passed correctly. I wanted to verify if this can happen with such a setup and any advise on debugging such a scenario?