Graph-based bootstrapped latent recommendation model

Abstract

As an important means to optimize organizational profitability, recommendation systems have been widely applied on e-commerce platforms in recent years. Their goal is to identify products of interest from which users have not browsed. To achieve this, prior work often relies on negative sampling strategies to guide the learning of user and product representations. In these strategies, products that users have not browsed are treated as negative labeled samples (products that users dislike). However, the negative sampling strategy fundamentally contradicts the goal of recommendation systems. With the number of products further increases, more “positive but not been browsed” products will be treated as negative labeled samples, leading to the introduction of noisy supervision signals during model training and thereby affecting recommendation performance. This paper proposes a Graph-based Bootstrapped Latent Recommendation model, dubbed GBLR. GBLR is a self-supervised framework that is trained using only positive user–product pairs. It utilizes a graph convolutional network to aggregate local neighborhood features of users and products, bootstrapping latent contrastive views. Subsequently, a symmetric cosine similarity loss function aligns the contrastive views of positive user-product pairs, guiding the model to learn consistent representations of users and products. With this self-supervised approach, the model can effectively learn the user and product representations in the absence of negative labeled samples. Experiments on three public datasets show that the proposed GBLR can effectively complete the recommendation task and outperforms the state-of-the-art baseline models. In the era of e-commerce, the innovative research on recommendation methods conducted in this work can optimize platform operations, enhance user experience and merchant revenue, thereby achieving a win–win situation for all parties involved, and holds significant practical value.