I currently run many of the ML models I keep through Shapley Additive Explanation (SHAP) values. I am curious about analyzing a raw dataset using SHAP values. Why shouldn't I be able to construct a raw passthrough model (X inputs -> y target) and run SHAP values on that? Would that not tell me what the most important factors are, of the factors/features I've chosen as X inputs, in the outcome of my y target? I am curious about using something like this as a baseline, and to compare trained XGBoost ML models (e.g. XGBoost) and their SHAP values with it.
Is there any fundamental reason why this does not make sense? I cannot seem to find anyone else mentioning this, but it makes sense to me why it should be useful. SHAP is a model-agnostic technique which just tells us how input factors influence a target outcome within a domain of observed values, using cooperative game theory. I don't understand why this is not used for descriptive statistics, given how popular it has become in machine learning. It seems like it could be an effective way of describing non-linear correlations between a target and other measured factors in the real world. And that in turn might make for an effective starting point when beginning subsequent modeling work (i.e. telling you the most predictive factors in a perfect model, AKA the real world over the domain of inputs and targets that have been measured).