Dual heterogeneous graph contrastive learning for QoS prediction

Abstract

The proliferation of Web Services leads to homogeneity issues, making accurate Quality of Service (QoS) prediction extremely helpful for inexperienced users to choose suitable services. However, the complex relationship between users and services in service invocation poses numerous challenges on QoS prediction. Given the capability of graph neural networks in modeling diverse relationships, a Dual Heterogeneous Graph Contrastive Learning method (DHGCL) is proposed in this paper to achieve high-accuracy QoS prediction. First, a dual heterogeneous graph is innovatively constructed, in which a global interaction graph is generated by a proposed graph learning to enable the direct interactions concerning the distant neighbors, while a local relationship graph is simultaneously constructed to enhance the close associations between users and services through spectral clustering. On this basis, the graph convolution network on the meta-paths is further designed to acquire the embedding of nodes for both of these two graphs. Finally, the global-local contrastive learning is served as a self-supervised mechanism to balance global interaction and local relationship information, and to complete the final QoS prediction. Extensive experiments have proven that our DHGCL method can achieve significantly higher accuracy than most of existing methods with the help of the dual heterogeneous graph.