Linear Variational Autoencoder for Top-N Recommendation
Zhou Pan, Wei Liu, Zaiqiao Meng, Jian Yin · 2022
Top-N recommendation is significant in various service-based platforms. Variational Autoencoders (VAEs) have been used in top-N recommendation in recent years for its effectiveness in collaborative filtering. Mult-VAE is such a variant that achieves great success, by adopting the multinomial likelihood, and an additional hyperparameter β on the KL divergence term of ELBO of VAE. However, Mult-VAE uses non-linear neural networks as encoder and decoder, to encode and reconstruct the user-item interaction data, which we prove unnecessary in sparse datasets because it will degrade the prediction accuracy in our experiments. Moreover, most variants of VAE-based collaborative filtering methods use the unnormalized user-item interaction data to make recommendations, which will hinder the learning process of the interaction data. In this paper, we propose Linear Variational Autoencoder (LVA), a linear version of Mult-VAE, considering additional normalization on user-item interaction data, for collaborative filtering under the implicit feedback setting. We verify its effectiveness in the experiments and prove that LVA achieves better or competitive performance over current state-of-the-art collaborative filtering methods, e.g, LightGCN, on four public real-world datasets.