Variational autoencoders

Variational autoencoders (VAEs) are a class of generative models that learn probabilistic latent variable representations of complex distributions. A VAE consists of an encoder (mapping data to a latent distribution) and a decoder (reconstructing data from latent samples), trained jointly by minimizing reconstruction loss and KL divergence (which regularizes the latent space towards a prior).

In the misinformation domain, VAEs are occasionally used for multimodal representation learning — learning a shared embedding space across text and image modalities by training the VAE to reconstruct both, encouraging the latent space to capture cross-modal correlations that a simpler concatenation or attention mechanism might miss. This is the key innovation of MVAE.

Key papers

• Khattar et al. (2019) — MVAE: First application of VAE to multimodal fake news detection on Twitter and Weibo. Learns shared text-image representations via joint training of encoder-decoder (VAE loss) and binary classifier. Achieves 74.5% / 82.4% accuracy on Twitter / Weibo, outperforming attention-based fusion baselines by ~6% accuracy.

Connections

• Multimodal fake news detection: VAEs are one architectural family for learning shared representations across modalities in news content.
• Deep learning methods in misinformation detection often use VAEs for unsupervised representation learning.