Compositional data are feature observations that describe amounts of components in relative portions that sum to a constant, typically 100%. Such feature data must be pre-processed before applying any machine learning predictive algorithm. They have been more frequently observed in insurance data so that methods of feature learning of such data are increasingly important. However, some cases of such multivariate observations have large number of zeros leading to sparsity. Traditional methods mapping them from simplex to real space may not be suitable when there is large sparsity. We explore methods within family of principal component analysis to handle such sparsity. Our numerical results using various datasets involving telematics and mining injuries produce superior predictive models with significant principal components. This is joint work with Guojun Gan and Banghee So.