Not sure how FM can be applied here. Generally used on user-content rating matrix kind of problems. We did think about using embeddings, but that would mean moving to a DL model which would increase inference time and we would loose out on interpretability aspect of decision trees.