On Improving Collaborative Filtering For The QoS Prediction Of Cloud Services

In recent years,the service resources in cloud center have been growing exponentially.It brings increasing difficulty for users to select high-quality services from a large number of candidates with equivalent functions.Quality of Service(Qo S)is an important indicator that users should pay attention to when they choose services to build a system.Unfortunately,in the real world,the Qo S history of many services in the candidate set is missing,because users can only invoke a small number of service resources in the cloud center.In order to obtain the missing Qo S data,the collaborative filtering(CF)technology in the recommendation system is applied to the fields of Qo S prediction,service selection and recommendation.However,the recommendation system has the inherent problems of sparse historical data and low prediction accuracy.To solve this problem,this paper proposes two novel methods to predict the missing Qo S values: For small-scale data sets,cloud service quality prediction with different filling methods based on different situations(DFDS)is proposed.Firstly,in order to solve the problem of very sparse historical data,a case-by-case pre-filling strategy is proposed,which divides the users into stable users and unstable users according to the stability.Similarly,services are divided into stable and unstable services based on stability.When combining user-based and service-based predictions,the proportion of the stable party should be appropriately increased.Secondly,the stability of users(or services)is considered when calculating the similarity between users(or services)by Pearson Correlation Coefficient.The similarity between users(or services)with stable consistency should be higher.In addition,the DFDS algorithm improves the average value of users(or services),that is,the dynamic weighted local average value is used to replace the static global average value.For large-scale data sets,cloud service quality prediction based on trend similarity using collaborative filtering(TSCF)is proposed.Firstly,in order to solve the problem that Pearson Correlation Coefficient cannot accurately calculate the similarity,a new similarity calculation method is proposed,which models the Qo S value of each user as a curve,so the problem of calculating the similarity between users is transformed into the problem of calculating the similarity between curves.Secondly,the weighted local average strategy in DFDS algorithm is still adopted in this algorithm.In addition,a post-processing strategy is proposed for abnormal Qo S predicted values(the predicted value of Qo S property such as response time and throughput is less than zero),which can automatically correct the prediction results of exceptions.Extensive experiments on real Qo S data show that the two Qo S prediction algorithms proposed in this paper are significantly superior to other popular Qo S prediction algorithms in prediction accuracy.