Sparse classification with positive-confidence data in high dimensions