ALCR: Adaptive loss based critic ranking toward variational autoencoders with multinomial likelihood and condition for collaborative filtering

Abstract

Research on variational autoencoders for collaborative filtering is gradually focusing on implicit feedback. However, most existing studies have two limitations: (1) they overlook the impact of user–item interaction data in implicit feedback on the representations of both users and items, which can affect the latent representations; (2) their attention is mainly focused on the immediate feedback of recommended items, ignoring interactions between feedback and ground-truth values, and neglecting the difference on loss functions between different training processes. To address these limitations, we first propose a condition for variational autoencoders to control user and item representations to learn more useful information from the latent representations. Then, we train an adaptive loss critic ranking to directly provide ranking scores in collaborative filtering recommendations, which aims to minimize loss and improve interactions during different critic training processes. Extensive experiments on three big real-world social media datasets demonstrate that this approach outperforms the existing twelve models under NDCG and Recall metric estimation settings and significantly improves the performance of a variety of prediction models.