A Two-Stage Approach for Learning a Sparse Model with Sharp Excess Risk Analysis