Quasiprobabilistic Density Ratio Estimation with a Reverse Engineered Classification Loss Function