Exploiting Web service geographical neighborhood for collaborative QoS prediction

Since Web services with equivalent functionalities but different quality are becoming increasingly available on the Internet, predicting the unknown QoS value of a Web service to an active user who has not accessed the service previously is often required for Web service recommendation and composition. Existing collaborative filtering methods suffer from the unavoidable sparsity and cold-start problems and underestimate the role of geographical information that inherently exists in user-service rating oriented model. The principal motivation for using geographical information in Web service QoS prediction stems from the observation that the ratings Web services perform are influenced significantly by their geographical neighborhood, a fact that is verified by our empirical data analysis on the real-world QoS dataset WSDream. Hence, it will be of interest to incorporate this implicit source of information in QoS prediction. In this paper, carefully selected geographical neighbors, clustered using a bottom-up hierarchical neighborhood clustering method, are smoothly integrated into a matrix factorization model, thereby building a more accurate prediction model. Further accuracy improvements are achieved by considering the biases of users and Web services. In experiments using the WSDream QoS dataset, our proposed method outperforms the other competitive methods with respect to accuracy and alleviates the sparsity and cold-start issues.