Instance hardness–based relevance for imbalanced regression

Abstract

Imbalanced regression problems occur when the target variable exhibits an asymmetric distribution, resulting in certain value ranges being underrepresented in the dataset. Traditional techniques for identifying these rare instances have limitations because they rely on the relevance function (ϕ) to define what is considered rare in the data distribution. This approach presents shortcomings in more complex scenarios, such as bimodal distributions, as it fails to adequately capture the notion of rarity by assigning fixed relevance values to all examples in the dataset, thereby compromising the distinction between truly rare and normal instances. Given these limitations, this study proposes an Instance Hardness-based relevance function, called the IHbased function, to identify rare instances in regression problems, as traditional relevance functions may fail in scenarios such as bimodal distributions, where rarity cannot be accurately inferred from the target values alone. By incorporating learning difficulty, the IH-based function offers a more reliable identification of truly rare instances.. Through experiments, we demonstrate that the IH-based function is able to correctly identify rare regions even in scenarios with bimodal distributions. Furthermore, the results show a significant improvement in model performance when using the IH-based function in balancing strategies such as Random Oversampling (RO) and Gaussian Noise (GN), compared to the traditional relevance function.