Research On QoS Prediction Via Network Biased Matrix Factorization

With the rapid development of the Internet,there exist a large number of Web services with the same or similar functions.In order to select the most suitable Web service from these candidates,users usually need to rely on some non-functional attributes,such as response time,throughput,failure rate,and reliability,which are often referred to as Quality of Service(QoS).If the QoS values of all services can be obtained,then the user can easily make a service selection.However,it is not easy to obtain the QoS value of all services,and there are three main challenges:1)The QoS value provided by each Web service is constantly changing according to its invocation object,and different users may observe very different QoS values for the same Web service.2)Considering the high time cost and huge resource overhead,it is impossible for users to invoke all Web services by themselves to obtain personalized QoS values.3)Some QoS attribute values are difficult to evaluate,and a small number of service invocations cannot calculate the real value of these attributes.Therefore,the above problems can be solved if the QoS values of Web services can be obtained in a predictive manner.In recent literature,Collaborative Factorization(CF)technology has been widely used for QoS prediction.The existing CF prediction methods can be divided into two categories:memory-based CF methods and model-based CF methods.The memorybased CF method identifies user neighbors with similar preferences based on their observed historical invocation records,which makes good use of local invocation information but lacks consideration of global invocation information.The model-based CF method builds a global model for QoS prediction based on observed historical invocation records,which focuses on calculating the interaction features between users and services,but lacks consideration of non-interaction features in the network environment.In this context,this paper proposes a QoS prediction method based on network bias and matrix factorization,and completes the following specific research work:1.A QoS prediction method based on network bias and matrix factorization(NBMF)is proposed,which first defines the QoS deviation caused by communication facilities and network environment as network bias,then uses linear regression method to predict the network bias between users and services,and uses matrix factorization method to capture the latent features of users and services during the interaction process,and finally combines the two methods linearly into the objective function of the prediction model by setting parameters.The NBMF approach takes into account both the interaction and non-interaction features between users and services,overcoming the limitations of existing QoS prediction techniques in adapting to diverse network environments and providing personalised QoS prediction results in both cold and hot start environments.2.This paper proposes a regularized NBMF method based on neighborhood relation(NBMF-R).NBMF-R first groups users and services respectively according to the network area information,then uses Pearson Correlation Coefficient(PCC)to calculate the similarity between users and services in the same network area,and finally integrates the regular terms based on neighborhood information into the NBMF method to minimize the potential differences between each user or service and its neighbors.The NBMF-R method makes full use of the hidden user neighborhood information and service neighborhood information in the userservice matrix,and can provide better prediction results than the NBMF method even in complex network environment.Finally,this paper conducts extensive experiments on the real world Web service data set.Firstly,the prediction performance of various methods is compared,and then the influence of various parameters of the proposed method on the prediction performance is analyzed.According to the experimental results,the proposed method can obtain better predictive performance than the existing methods in the same context.