Reproduce “RecVAE: a New Variational Autoencoder for Top-N Recommendations with Implicit Feedback” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ilya Shenbin, Anton Alekseev, Elena Tutubalina, Valentin Malykh, Sergey I. Nikolenko
RecVAE, developed at the Samsung-PDMI Joint AI Center, improves top-N recommendations with implicit feedback by introducing a denoising variational autoencoder with a novel encoder architecture, a composite prior distribution, user-specific KL divergence scaling, and an alternating training strategy, outperforming Mult-VAE and other autoencoder models on benchmark datasets.
Key insights
- Stabilizing VAE training, especially to mitigate "forgetting" in sparse data, can be achieved by regularizing large variational parameter updates with a composite prior that incorporates past knowledge.
- Optimizing VAE regularization for implicit feedback benefits from user-specific KL divergence weighting, where the regularization strength adapts to the density of a user's feedback.
- An alternating training strategy, where a complex encoder is trained with denoising and a simpler decoder is trained on clean inputs without KL regularization, can significantly improve performance by avoiding over-regularization and balancing component complexities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.