Research On Web Service QoS Prediction Method Based On Variational Autoencoder And Denoising Autoencoder

Users must sift through the huge selection of network services in the Internet age to choose the one that best suits their needs,which is like to tring to find a needle in a sea.And personalized Web service recommendation technology and methods can solve this problem.However,the proliferation of mobile devices has led to an increasing number of services with similar or identical functionality,affecting the user experience.So the response time and reliability of services are what users are more bothered about.Quality of Service(QoS),a quality that gauges a Web service’s non-functional performance,is an essential indicator for personalised service recommendation.However,existing methods and techniques based on QoS prediction still need to address some unavoidable problems,such as sparse raw data,the influence of noisy data,lack of consideration of user and service-related information,inadequate extraction of potential features from QoS data,cold starts and model overfitting.In current research,the above problems persist.This thesis studies the QoS prediction method of Web service,and its primary research contents include the following:(1)To improve the precision of QoS prediction under the condition of sparse original data and noisy data,as well as the potential correlation information between users and services,this thesis proposes a QoS prediction model based on multiple double-layer stacked noise-reducing autoencoder.Firstly,considering that the geographical location information of users and services has a certain influence on QoS,and then simply dividing users and services based on location information.The traditional Jaccard similarity calculation method is improved to calculate and obtain similar neighbours that are in the same location area as the target user or service,so as to pre-populate the original sparse QoS matrix and alleviate the sparsity problem.According to the frequency that each service is called with similar neighbors,the user’s preference information of service invocation is captured and combined with the prepopulated QoS matrix.The processed QoS data is then subjected to multiple noise addition and reconstruction by a multiple double-layer stacked noise reduction autoencoder model to fully exploit the potential QoS data features and improve the QoS prediction accuracy.(2)Further considering the problems of cold start and model overfitting,this thesis proposes a hybrid model for QoS prediction based on an improved conditional variational autoencoder.To solve the data sparsity problem by pre-populating the missing values of the sparse QoS vector,the Bias based Singular Value Decomposition(Bias SVD)is chosen.The pre-populated QoS vector is then downscaled using a twolayer stacked autoencoder to extract potential features of the user as external conditions for the model.The external conditions and the pre-populated QoS vector are learned through training of the conditional variational autoencoder to reconstruct the missing QoS values.Based on the location information,the Pearson Correlation Coefficient and Independent Samples T-test are employed to calculate similar neighbors who have excellent reputation.They are combined with the reconstructed QoS matrix for prediction to further improve the QoS prediction accuracy.At last,the WSDream data set is used for testing the two prediction models.As well as the outcomes demonstrate the viability of the QoS prediction model based on multiple double-layer stacked noise-reducing autoencoder.The prediction accuracy of hybrid QoS prediction model based on improved conditional variational autoencoder is increased.