Invertible TabMap: An Invertible Self-supervised Mapping for Imbalanced Classification of Tabular Data

Classification with imbalanced datasets has garnered escalating attention as it is inherent in various applications. In the context of imbalanced tabular data classification, columns may have multiple and skewed distributions, making existing methods fail to properly model this data structure. To address this problem, we propose the Invertible TabMap, a novel framework creatively amalgamating a newly devised self-supervised paradigm named TabMap with a designed invertible network. Different from existing self-supervised learning for tabular data, TabMap maximizes the mutual information between global and specifically designed local structures. These local structures are extracted from specific feature combinations that reveal the distinctive essence between imbalanced classes, and they are robust to permutations of tabular columns. TabMap aims to create a latent space that aligns with these abundant and diverse local structures, from which effective representations are generated for classification. To further mitigate data imbalance to train a better classifier, we formulate an invertible network capable of simultaneously generating minority class samples and selecting high-quality samples. The generation process generates minority class samples in the TabMap latent space by an invertible generator. Subsequently, the selection module operates in the invertible latent space, selecting high-quality samples to train classifiers. Extensive experiments demonstrate that the proposed framework leads to significant improvements in the acquisition and synthesis of imbalance information, and it exhibits state-of-the-art performance for the classification task of imbalanced tabular data compared with other existing alternative methods.