Frequency-aware priors for variational autoencoders under class imbalance

초록

Variational Autoencoders (VAEs) often yield distorted latent representations under class imbalance, as minority classes collapse due to excessive regularization from the isotropic Gaussian prior. We propose Frequency-Aware Prior VAE (FAP-VAE), which adapts the prior by scaling variance inversely with class frequency and by updating both mean and variance via exponential moving averages of encoder-derived statistics. This design alleviates over-regularization of minority classes and preserves latent separability. Experiments on seven benchmarks show that FAP-VAE improves clustering quality—achieving higher normalized mutual information (NMI) and adjusted Rand index (ARI)—while maintaining reconstruction and generation performance. These results indicate that data-aware prior adaptation enhances the robustness of representation learning under imbalanced conditions. © 2026

키워드