Meta learning-based relevant user identification and aggregation for cold-start recommendation

Cold-start has always been a major concern for recommendation systems. Heterogeneous Information Networks (HINs) are widely used due to their ability to provide rich auxiliary information to enhance the representation of users or items, effectively mitigating the cold start problem. However, the rich semantic information provided by relevant users is often overlooked. Meanwhile, the scarcity of training data has made meta learning widely used for cold start recommendation since it can learn general knowledge from a small amount of data and quickly adapt to new tasks. To address this issue, we propose a cold-start recommendation model based on meta learning for relevant user identification and aggregation, called IAML. First, IAML identifies relevant users with similar preferences by integrating meta-path-based and clustering methods, and constructs them into a neighbor set. Next, the information from neighbor nodes and their interactions are integrated to assess their impact on cold-start nodes, thereby obtaining a richer user representation. In view of the data sparsity in the training process, an optimized meta-learning algorithm MAML is introduced to enhance the model’s generalization ability with a limited amount of training data. Finally, extensive experimental results on three public datasets show that our IAML exhibits satisfactory performance across all metrics, both in cold-start and non-cold-start scenarios.