Adaptive Graph Neural Networks for Cold-Start Multimedia Recommendation

In recommendation systems, the user cold-start problem remains a critical challenge due to sparse user-item interactions. Recently, emerging research efforts have been devoted to leveraging Graph Neural Networks (GNNs) for recommendation systems, yielding promising performance. However, few of them investigate the natural mechanism of GNNs, i.e., which neighbors to include and how to aggregate emerging multimedia information, to solve data sparsity problems. It is important to determine user-specific propagation and aggregation mechanisms, called receptive fields, of GNNs specifically for recommendations to estimate cold-start users' preferences with sparse user-item interactions. Therefore, in this paper, we propose a novel REceptive fields based ADaptive Graph Neural Network for user cold-start recommendation, named READ. The key insights of the tailored receptive fields in READ are two-fold: A sparsity receptive aggregation mechanism is carefully designed, integrating user interaction sparsity into the aggregation's breadth and depth; furthermore, a sociality receptive aggregation mechanism is brought up to make full use of social side information from multimedia content, particularly high-order social relations derived from the interaction history, to replenish the estimation of user interests and preferences. Extensive experimental results on two real-world industrial datasets and one benchmark dataset show that our proposed READ significantly outperforms the baselines.