Research On QoS Prediction Approach With Latent Feature

With the rapid development of Internet technology,the number of web services is increasing rapidly,and there are many web services with the same or similar functions in the network,so it is difficult for users to choose the services that meet their own needs.In order to effectively solve this problem,service recommendation has attracted extensive attention from researchers,especially the recommendation task based on Quality of Service(QoS).QoS is defined as a set of non-functional user experience properties,it is the key criterion of service recommendation and service selection,which can distinguish the suitable services among different functionally equivalent services.However,it is really hard for a user to invoke all candidate services to acquire all QoS values to make a final decision,so QoS prediction of web services is an essential step for high-quality service recommendation.In recent years,some QoS prediction methods have been proposed,among which Collaborative filtering(CF)has become the main prediction method because of its simplicity and efficiency.Collaborative filtering methods can be divided into two types: neighbor-based CF and model-based CF.The neighbor-based CF has a strong ability to learn local features,but it cannot effectively solve the problems of sparse and abnormal data.Model-based CF has a strong ability to learn global features,and can still provide good prediction results in the case of sparse data,but it lacks the ability to learn local features and cannot learn the nonlinear features in the data.In view of the appeal situation,this paper proposes two new prediction methods by combining the advantages of neighbor-based CF,model-based CF and deeplearning-based prediction methods: the improved autoencoder based hybrid prediction approach and cluster feature based context-aware neural collaborative filtering.The two methods combine traditional collaborative filtering method and deep learning method,which can use the deep learning model to learn deep nonlinear latent feature in QoS and alleviate the problem of data sparsity.The main contributions of this paper are as follows:(1)This paper proposes an Improved Auto Encoder based Hybrid prediction approach(IAEH).IAEH is a hybrid prediction method that combins improved Auto Encoder and neighbor-based CF.The prediction process of IAEH can be divided into two phases: global feature learning and local feature learning.In the global feature learning phases,this paper proposes an improved Auto Encoder,which alleviates the data sparsity problem by filling in the sparse input vector,and then uses the improved Auto Encoder to learn the nonlinear latent features in QoS data.In the phase of local feature learning,a new method of similarity calculation is proposed to alleviate the problem of overestimation of neighbors by introducing two factors: the common invocation factor and the invocation frequency factor.(2)This paper proposes a cluster feature based Context-aware Neural Collaborative Filtering(CNCF).CNCF is a prediction method combining Possibilistic fuzzy Clustering Means(PCM)and Neural Collaborative Filtering(NCF).Its prediction process includes two steps.First,context information clustering.This paper uses PCM algorithm and context information to cluster users and services.Because context information contains a variety of property data,this paper proposes a distance calculation method that can comprehensively analyze different properties.Second,deep latent feature learning.In this paper,an improved NCF model is proposed,which can learn cluster features from the contextual information and deep latent features from historical QoS records at the same time.(3)The WSDream dataset is used to verify the prediction method in this paper.This dataset is widely used in QoS prediction research,and has response time and throughput two QoS data.A variety of sparse environments in response time and throughput data are verified.After a large number of offline experiments,the analysis of the experimental results shows that the IAEH and CNCF methods proposed in this paper have high prediction accuracy.The method proposed in this paper can partially solve such problems in collaborative filtering,such as low prediction accuracy when data is sparse,inability to extract deep nonlinear latent features,insufficient use of context information,etc.The method proposed in this pape can promote the research on QoS prediction by using latent features.