CARTGen-IR: Synthetic Tabular Data Generation for Imbalanced Regression

TL;DR

A CART-based synthetic sampling method specifically designed for imbalanced regression on tabular data that integrates relevance- and density-guided sampling to address sparse target regions without thresholding, and employs a feature-driven tree structure to generate realistic tabular samples across heterogeneous features and non-linear interactions.

Abstract

Handling imbalanced target distributions in regression poses a persistent challenge, as the underrepresentation of relevant target values can significantly hinder model performance. Existing data-level solutions often adapt classification-oriented techniques, introducing arbitrary thresholds over the continuous target and leading to artificial and potentially misleading problem formulations. Deep generative models offer flexible sample synthesis but are computationally intensive and difficult to interpret. We propose a CART-based synthetic sampling method specifically designed for imbalanced regression on tabular data. The method integrates relevance- and density-guided sampling to address sparse target regions without thresholding, and employs a feature-driven tree structure to generate realistic tabular samples across heterogeneous features and non-linear interactions. Experiments on benchmark datasets for extreme-value prediction show that the proposed approach is competitive with state-of-the-art resampling and generative methods while offering faster execution and greater transparency. These results highlight its potential as a scalable and interpretable data-level strategy for improving regression models in imbalanced domains.

Figures (6)

• Figure 1: CART consecutive column-wise data generation.
• Figure 2: Synthetic sample generation comparison.
• Figure 3: Comparison of wins and losses, including significant outcomes at the 95% confidence level, for each data-level strategy against a no-preprocessing baseline, across different learners, datasets, and evaluation metrics.
• Figure 4: Bayesian posterior ternary plots.
• Figure 5: Sensitivity analysis of CARTGen-IR with respect to the density weighting scheme ($\rho$), exponent ($\alpha$), sampling proportion ($\eta$), and noise level ($\delta$).