More data speeds up training time in learning halfspaces over sparse vectors