Abstract:The collaborative recommendation algorithm based on the Variational Autoencoder (VAE) can help solve the sparsity problem in the recommendation algorithm, but the VAE model’s prior is a single Gaussian distribution, which makes the expression tends to be simple and average, and suffers from the problem of underfitting. The Gaussian Mixture Variational Autoencoder (GMVAE) model has a more complex prior, which is more adaptable and effective for nonlinear tasks compared to the original VAE model, and has been widely used for unsupervised clustering and semisupervised learning. Inspired by this, this paper investigates a collaborative filtering algorithm based on the GMVAE model. In this paper, the authors design experiments based on the Cornac recommender system comparison framework, and use the improved GMVAE for the collaborative recommendation task, the useritem matrix regenerated by the generative model is used for recommendation task. Deep features are extracted with one hidden layer in the inference model and one layer in the generation model to increase model robustness, and an early stop strategy is used to reduce overfitting. In this paper, experiments are conducted on multiple public datasets to compare with other recommendation algorithms in terms of NDCG and recall metrics. The experiments demonstrate that the improved collaborative filtering algorithm based on a GMVAE model performs well in the recommendation task.