Meta-Path Guided Heterogeneous Social Occasion Service Composition Embedding Based on BERT

Social occasion service is a popular service type which is available in many lifestyle apps. Package style value-added social occasion services can be provided to achieve more complex functions for requesters needs by service composition. But with the increase in the types and quantities of atomic services, commonly used service composition processes face difficulties in service decomposition, requiring consumers to have more prior knowledge, clearer intentions, and more manual operations. Therefore, automatic composition and discovery of social occasion services have emerged as a crucial potential area, and finding an embedding learning method for social occasion service composition is a key step in realizing it. In this study, we transform the embedding problem of social occasion service composition into a representation learning problem of node compose in a heterogeneous graph firstly. Then, meta-path guidance idea which has been proven to be effective on heterogeneous graphs, is introduced for generating a large number of feasible service compositions and multi-class training guidance sequences. At last, the BERT model with special classification tasks is used to learn the generated service sequences, for extracting the features and forming the embedding of service compositions. The experiment shows that our approach can learn social occasion service composition well and that the representative vectors generated by the pretrained model can be utilized to help implement related service discovery.