Modern compositional datasets in many fields (e.g. microbiome research and econometrics) display traits such as high-dimensionality and sparsity that are poorly modelled when using traditional log ratio-based models. While the performance of machine learning methods for in-sample prediction on such datasets remains impressive, their ability to generalise beyond the convex hull of the data points is limited. Further, the black box nature of the algorithms necessitate additional procedures to estimate the (causal) effect of individual features on outcomes of interest. In this talk, we present semiparametric efficient estimators of two model-free variable importance measures for compositional predictors. The first measure is tailored to quantifying the effect of a frequently observed target feature relative to the remaining features and the second measure compares the effect of the target feature being zero and non-zero. Our proposals are based on decoupling the target feature from the remaining features by projecting the remainder into a lower-dimensional simplex, thus making the target and remainder variationally independent. We provide conditions under which our estimators are asymptotically Gaussian achieving the efficient variance bound. We demonstrate the performance of our proposed methods through extensive simulation and real data experiments.