Towards imbalanced regression over distributionally biased data: A fast static approach

The generalization of models are susceptible to data bias for both classification and regression problems, which also has intrinsic connection to the issues of fairness. However, existing approaches focus on class-imbalanced learning and fail to address these concerns for regressors. In this paper, we target at imbalanced regression with particular focus on fusing distributional information from both the feature space and target space. We first introduce two metrics, uniqueness and abnormality, to reflect local data distribution and assess the informativeness of each sample from a regional perspective in the two spaces. By integrating these two metrics we propose a local Variation-incented re-weighting method, termed ViLoss, which fuses distributional information for each sample to optimize gradient-based regressor training. The weights are computed once-and-for-all in pre-processing and thus our method causes little extra computation during training. Empirically, we conducted comprehensive experiments on both synthetic and real-world datasets for parameter study and performance evaluation. The results demonstrate the efficacy of our method in boosting model quality (error reduction by up to 39.0%) as well as narrowing the gap of error between groups.