A service-recommendation method for the Internet of Things leveraging implicit social relationships

Abstract

Integrating social-relationship information is widely considered an effective means to addressing the sparsity issue in Internet-of-Things (IoT) service recommendations. However, owing to platform variations and privacy concerns, acquiring explicit social-relationship information has become challenging. Therefore, researchers are gradually focusing on leveraging implicit social-relationship information to enhance recommendation effectiveness. Nevertheless, when using implicit social relationships, both user/service-implicit social-relationship and user-service interaction information rely on the same rating matrix, leading to nonindependence and coupling, which influence the recommendation model. To address this challenge, the present paper introduces a unique approach for IoT service recommendations, leveraging implicit social-relationship information (short for ISoc-IoTRec). First, we construct a user-service interaction graph, user-implicit social-relationship graph, and service-implicit social-relationship graph, learning their node embeddings through graph neural networks (GNNs). Subsequently, we introduce a cross information control module to achieve feature separation, ensuring that the user and service embeddings learned from different graphs remain independent in representation, thereby alleviating the nonindependence and coupling issues arising from the same data source. Following feature separation, the user and service embeddings are aggregated separately. Through an attention mechanism module, the model can selectively emphasize or attenuate the impact of each feature while considering the overall information, further addressing nonindependence and coupling issues. Extensive experiments conducted on three real-world datasets underscore the remarkable performance of ISoc-IoTRec, significantly outperforming existing recommendation algorithms.