Low-rank approximations for predicting voting behaviour