Regularization for Deep Imbalanced Regression Based on Quantitative Relationship

Imbalanced datasets are prevalent in real life. The imbalanced datasets pose challenges for classification and regression tasks. Compared to imbalanced classification, imbalanced regression deals with continuous labels. The positional relationship of the labels is unable to reflect the relationship of the feature space. The feature of the rare sample is under-represented due to the influence of frequent samples. The Quantitative Relationship (QuanRel) regularizer is proposed to mitigate the problem of under-representation of features. The effect of frequent samples on the rare samples in the feature space is mitigated by the QuanRel. The QuanRel uses the number of discriminated classes, instead of the information in the label space. The incorrect proximity of features is improved by the QuanRel. Three regression benchmark datasets are used to demonstrate the effectiveness of QuanRel. The results of the models with QuanRel are the best of the common models in the imbalanced regression datasets.