Towards the compression of parton densities through machine learning algorithms