Ponderação baseada em expertise para modelos de regressão com rótulos ruidosos

Abstract

Traditional regression methods assume the availability of precise labels for training models. However, in many contexts, obtaining fully accurate labels may not be feasible, requiring reliance on multiple experts whose opinions may diverge due to intrinsic human noise, which is difficult to measure. This noise can be present in the input variables, as different experts may interpret certain observations in distinct ways due to their expertise. In this work, we propose an innovative approach to training regression models in scenarios in which the labels contain noise, resulting from multiple divergent expert opinions. The proposed method first estimates each expert’s expertise both generally and at the instance level, assigning weights to their opinions. Then, a weighted average of these opinions is computed, using the learned weights to adjust the regression model based on the input variables. The proposed approach has a solid theoretical foundation and, through experiments with both simulated and real data, has been empirically demonstrated to outperform traditional methods. In summary, this method provides a simple, fast, and effective solution for training regression models in scenarios with noisy labels generated by differing expert opinions.