HardTVAE: Hardness-Aware Generation and Multi-View Fidelity Evaluation of Synthetic Tabular Data for Imbalanced Learning

Abstract

We introduce HardTVAE, a hardness-aware tabular variational autoencoder designed to improve synthetic data generation in the presence class imbalance. HardTVAE modulates the evidence lower bound with instance-level hardness weights implemented via static, curriculum, or self-paced strategies, and remains agnostic to the choice among the seventeen supported hardness metrics. Complementing the generative component, we propose a multi-view fidelity framework that integrates distributional, topological, complexity-based, and hardness-based perspectives to capture complementary aspects of data fidelity and reveal structural, geometric, and instance-level properties beyond what any single view captures. The multi-view framework is operationalised by fusing the four fidelity views using the harmonic mean to form the Multi-View Fidelity Index (MFI). We evaluate our approach on ten real-world healthcare datasets, a domain where pronounced class imbalance provides a demanding testbed, and validate all comparisons with non-parametric statistical testing. HardTVAE establishes a new performance trade-off compared to both baselines. It significantly exceeds TVAE in downstream utility while still frequently outranking it in fidelity, and it significantly exceeds CTGAN in multi-view fidelity while remaining ahead on average in utility. This advance is realised most consistently by hardness measures such as Class Likelihood (CL), Collective Feature Efficiency (F4), Maximum Individual Feature Efficiency (F3), Tree Depth Unpruned (TDU), Ratio Intra/Extra Class Distance (N2), whose configurations rank above both baselines on fidelity while securing significant utility gains. The proposed generative model and evaluation framework establish a structured basis for both enhancing and evaluating synthetic data under class imbalance.