# RecVAE: a New Variational Autoencoder for Top-N Recommendations with Implicit Feedback

**Authors:** Ilya Shenbin, Anton Alekseev, Elena Tutubalina, Valentin Malykh, Sergey I. Nikolenko  
**arXiv:** [1912.11160](https://arxiv.org/abs/1912.11160)  
**Date:** 2019-12-24  
**Summary page:** https://www.summarizepaper.com/en/paper/1912.11160v1/

## Key points

- Recent research has shown the effectiveness of using autoencoders based on deep neural networks for collaborative filtering.
- Mult-VAE is a notable model that utilizes multinomial likelihood variational autoencoders and performs exceptionally well in top-N recommendations.
- RecVAE is a new model that builds upon Mult-VAE and introduces several key improvements.
- It introduces a novel composite prior distribution for the latent codes to improve modeling of user preferences.
- It presents a new approach to setting the β hyperparameter within the β-VAE framework for better control over reconstruction accuracy and regularization trade-off.
- It adopts a new training strategy based on alternating updates to enhance performance.
- Extensive experimental evaluations across classical collaborative filtering datasets demonstrate that RecVAE significantly outperforms previous autoencoder-based models like Mult-VAE and RaCT.
- A detailed ablation study is presented to provide a comprehensive assessment of the developments.
- Code and models for RecVAE are available on GitHub for reproducibility and further exploration of its capabilities.

## Summary

Recent research has demonstrated the effectiveness of using autoencoders based on deep neural networks for collaborative filtering. One notable model is Mult-VAE, which utilizes multinomial likelihood variational autoencoders and has shown exceptional performance in top-N recommendations. Building upon this work, the authors propose a new model called RecVAE (Recommender VAE) that incorporates novel ideas to further enhance Mult-VAE. RecVAE introduces several key improvements over Mult-VAE. Firstly, it introduces a novel composite prior distribution for the latent codes which helps improve the modeling of user preferences. Additionally, RecVAE presents a new approach to setting the $\beta$ hyperparameter within the $\beta$-VAE framework allowing for better control over the trade-off between reconstruction accuracy and regularization. Lastly, RecVAE adopts a new training strategy based on alternating updates which further enhances its performance. To evaluate the effectiveness of RecVAE, extensive experimental evaluations are conducted across classical collaborative filtering datasets. The results demonstrate that RecVAE significantly outperforms previously proposed autoencoder-based models such as Mult-VAE and RaCT. To provide a comprehensive assessment of their developments, the authors also present a detailed ablation study. The availability of code and models on GitHub (https://github.com/ilya-shenbin/RecVAE) allows for reproducibility and further exploration of RecVAEs capabilities. Overall, this research contributes to advancing autoencoder-based approaches for collaborative filtering by introducing RecVAE—a model that surpasses existing methods in terms of recommendation quality.