GraViti: Graph-Level Variational Autoencoders with Relaxed Permutation Invariance
For researchers working with graph-structured data (e.g., molecules, Bayesian networks), GraViti provides a lightweight, single-step generative model that outperforms existing methods in reconstruction while offering smooth interpolation and property-guided search.
GraViti introduces a transformer-based graph-level VAE that maps entire graphs to compact latent vectors, achieving state-of-the-art reconstruction accuracy on large molecular datasets and demonstrating that enforcing permutation invariance can be detrimental for consistent reconstruction when a canonical node ordering exists.
We introduce GraViti, a transformer-based graph-level variational autoencoder that maps entire graphs to compact latent vectors. This design produces a true graph-level latent space that supports smooth interpolation, property-guided search, and other downstream tasks beyond the constraints of node-level embeddings. On molecular benchmarks, GraViti learns to decode valid samples that follow the chemical constraints present in the training data, showing that the model recovers domain rules directly from graph-level representations. We also show that, in domains where a reliable canonical node ordering exists such as molecules or bayesian networks, enforcing permutation invariance can prove detrimental for consistent reconstruction. GraViti achieves state-of-the-art reconstruction accuracy on large datasets, and provides solid generative performance. Its single-step decoding offers a lightweight alternative to more complex generation pipelines while maintaining practical sample quality.