This paper builds on numerous recent works that focus on defining interpolation operators for algebraic multigrid in the case where traditional approaches fail due to overly restrictive assumptions. Here, a method is proposed that falls within the “bootstrap” variety of these approaches, where each row of interpolation is defined by a least-squares fitting process. The particular novelty is in the addition of a residual term to the fit that penalizes the fit in the case of large local residual values for the error being fit. A simplified analysis is provided that suggests the approach is better than others recently proposed. A motivating example from lattice quantum chromodynamics provides the bulk of the numerical results.