Reducing high-dimensional data