Toward Deep Autoencoder for Recommendation System Using Implicit Feedback
Ajoy Deb Nath, Md. Mokammel Haque · 2025
Recent years have seen an exponential increase in the use of recommendation systems. It is necessary to filter, prioritize, and effectively convey relevant information to consumers in order to suit their tastes and preferences on the online platform, where the sheer volume of options is available. In this chapter, we present a recommendation model that yields top-N recommendations based on collaborative filtering. We strengthen the Variational Autoencoder (VAE) by incorporating residual connections in encoder architecture, improvising mixture prior distribution using hierarchical priors , and introducing β divergence as the objective function. Empirically, we analyze the impact of β on the objective function. Through a number of experiments, we also emphasise the significance of latent space in the model. By considering implicit feedback, we construct a user-item interaction matrix and feed the matrix to our autoencoder model. In our investigation, we use two datasets, and the results show that the MovieLens-20M dataset produces the best outcome, while the Netflix Prize dataset provides competing results. On the ML-20 evaluation dataset, our proposed model receives a Normalized Discounted Cumulative Gain ( nDCG ) score of 0.49, indicating the effectiveness of our method in capturing relevant recommendations within the top 100 ranked items.