Research Of QoS-Based Web Service Recommendation

In Qo S-based Web service recommendation system, predicting Quality of Service(Qo S) for users will greatly aid Web service selection and discovery. Collaborative filtering(CF) algorithm is an effective method for predicting Web services’ Qo S values. Collaborative filtering algorithms could be divided into two main categories: memory-based and model-based algorithms. Memorybased CF algorithms makes prediction by similar users or similar services. The prediction accuracy of memory-based CF algorithm is pretty high. But they suffer from potential scalability and data sparsity issues. Model-based CF algorithm makes prediction by the learned model which has been acquired by machine learning or data mining algorithms. Although model-based CF algorithm can address scalability and data sparsity problems, but the prediction accuracy of those algorithms is rather lower and the model employed by model-based CF algorithm is time-consuming to build and update. The state-of-the-art CF algorithms, both of memory-based CF algorithms and modelbased CF algorithms, aren’t fit for Web service-oriented application environment, since the Web service-oriented application environment is a high dynamic and larger-scale environment.This paper focus on improving the prediction accuracy of CF algorithms. Further, this paper not only address scalability and data sparsity problems, but also lower the time complexity of model building and updating.The key contributions of our work is as follows:Improving the prediction accuracy. In order to improve the prediction accuracy, three different factors are taken into account when the similarity of users or Web services are calculated and the missing Qo S values are predicted. The factors which have been employed by this paper, are time, load, user input and the time complexity of Web service. This paper take the time complexity of the Web service and the user input as a unified whole and integrates the time complexity of the Web service and the data size of user input into the similarity measurement and the Qo S prediction. To the best of our knowledge, none of the existing CF approaches takes time or load into account. And none of the existing CF approaches takes the time complexity of Web service into account, not to mention that the time complexity of Web service should be considered together with the data size of user input as a unified whole. Besides, the influence weights have also been employed to automatically combine the enhanced user-based and item-based methods to improve the prediction accuracy. Generally, a tunable parameter λ has been employed to combine the enhanced user-based and item-based methods. The value of tunable parameter λ would impact the prediction accuracy significantly. And it is very difficult to find the optimal values of tunable parameter λ.Thus, the influence weights are simpler and more practical than the tunable parameter λ.Addressing the scalability problem. Network distance has been employed to group users or Web services. Then it seeks users or Web services for recommendation within smaller and highly similar clusters instead of the entire database. Both geographic distance and AS have been employed to measure network distance. The advantage of this approach is that both user clusters and Web service clusters are constructed and updated quickly.Alleviating the data sparsity problem. To alleviate the data sparsity problem, Firstly, users or Web services would be grouped by traditional classify algorithm, such as K-Means, or network distance. Secondly, converting user-service matrix into user Cluster-service matrix and user-service Cluster matrix to improve matrix density, since the number of user clusters or Web service clusters is usually much smaller than the number of users or Web services.Finally, the user Cluster-service matrix is used to calculate the similarity of Web services and the user-service Cluster matrix is used to calculate the similarity between users. According to the experimental results, it would improve the prediction accuracy obviously when the data density of user-service matrix is sparse.Acquiring the testing data set. Since the real state-of-the-art Qo S data sets of Web services don’t include load, user input and the time complexity of Web service, in order to evaluate the approach proposed by this paper, dozens of Web services had been developed and deployed in 7 different regions of the Amazon E2 C to simulate Web services coming from different locations. Users had been deployed in 4 different cities of China. Thus, the realworld Qo S values have been collected from these users. Series of real-world experiments have been conducted based on those real-world Qo S values.