Research On QoS Prediction Based On Deep Learning And Matrix Factorization

The advances in eCommerce,especially the pay-as-you-go cloud model have fuelled the rapid growth of Web services.The statistics published by Programmable Web,an online Web service repository,indicate a rapid growth in the number of published Web services over the past few years.The popularity of Web services and service-oriented architecture(SOA)allows different service-oriented applications to be built to fulfill various organizations',increasingly sophisticated business needs.However,with the rapid growth of the number of Web services,a large number of services with same functionality have emerged,how to choose the best service for users from those functionally equivalent candidate services has become a challenging research topic.Quality of Service(QoS)is widely used to describe and evaluate non-functional attributes of Web services.Currently,QoS is successfully applied to service recommendation.In the Web service recommendation,it is necessary to predict the missing QoS value of services,and the application of Matrix Factorization(MF)technology has attracted people's attention.Recently,researchers have begun to use user similarity to improve MF-based Web service prediction methods.However,the existing methods do not solve the two main problems very well and systematically:1)take appropriate neighborhood information;2)make full use of neighborhood information.In addition,the matrix factorization is applied to the inner product of the potential features of users and services.But inner product is only a linear combination of potential features and may not be sufficient to capture the complex structure of user interaction data.Therefore,this paper proposed two novel methods for predicting quality of service.The main contributions of this paper are as follows:(1)A covering-based quality of service prediction method is proposed(CNMF),which is a covering-based quality prediction method for Web services via neighborhood-aware matrix factorization.The novelty of CNMF has two aspects.First,it used a covering-based clustering approach to find similar users and similar services.The clustering method does not require pre-specified number of clusters and cluster centers.Second,the neighborhood information of users and services is used to improve prediction accuracy.CNMF first employed the covering algorithm to calculate the similarity between users and users,services and services;then it used Top-k mechanism to select similar users and similar services;finally,intergrated user neighborhood information and service neighborhood information into matrix factorization model to predict QoS.The experimental results show that CNMF significantly outperforms eight existing quality prediction methods,including two state-of-the-art methods that also utilize neighborhood information with MF.(2)A quality prediction method for Web services via deep learning is proposed(DeepL).The method first obtained user potential vector and service potential vector by using user ID and service ID,then used the user potential vector and the service potential vector as input of the deep learning model.To make the DeepL method nonlinear,we used a multilayer perceptron to learn the nonlinear relationship between users and services.The experimental results show that the DeepL method is superior to the existing methods in predicting the QoS missing value by proposing the nonlinear relationship between user and service,including the current mainstream matrix factorization method.(3)The effectiveness of the proposed methods are verified by experiments using a large-scale Web service QoS dataset in a public real environment,which contains a total of 1,974,675 quality records for 5,825 web services generated by calls from 339 users distributed around the world.Each record has two values,including response time(RT)and throughput(TP),which is the largest QoS dataset in the real-world environment that has been publicly released.