When building ML systems in domains with expert knowledge, use SCCR to audit whether your raw-input model is learning something genuinely new or just redundantly encoding expert features—this helps identify where expert views fall short.
This paper introduces SCCR, a method to understand what predictive information raw-input models capture beyond expert-designed features. By selectively removing shared directions between raw and expert embeddings, the framework reveals gaps in expert views while preserving unique signals, tested on medical imaging and clinical prediction tasks.