In this work we improve the existing tools for the recovery and prediction of human decisions based on multiple factors. We use essentially a latent factor method, and obtain the decision-influencing factors from the observed correlations in the available statistical information by singular value decomposition-based principal factor identification. We generalize on widely-used linear representations of decision-making functions by using adaptive high-order polynomial interpolation and applying an iterative and adaptive post-processing to arrive at an estimated probability of every possible outcome of a decision. The novelty of the method consists in the use of flexible, nonlinear predictive functions, and in the suggested post-processing procedure. Our experiments show that the introduced approach is at least competitive in the class of SVD-based prediction methods, and that the precision grows with the increase in the order of the polynomial basis. We suggest that the method may be successfully applied instead of a widely used linear SVD-based methods.